Memnest Memory MCP Server

Persistent graph memory for AI agents using LadybugDB — an embedded graph database with native vector search and full-text search.

Give your AI agent memory that persists across sessions, deduplicates automatically, and models knowledge as a graph with typed relationships.

Why Memnest?

• Graph memory — memories linked via Topic nodes and relationships (RELATED_TO, SUPERSEDES, EXPLAINS) with Cypher queries
• Three-layer auto-dedup — exact hash + semantic similarity + LLM-driven consolidation
• Workspace namespacing — memories scoped per project; global_search opt-out
• HNSW vector search — fast cosine similarity over FastEmbed embeddings
• Topic auto-linking — tags become graph nodes, enabling traversal queries
• Embedded — no Docker, no server process, single database directory
• Zero config — sensible defaults, just install and run
• Importance & access tracking — memories ranked by relevance and usage

Benchmarks

Memnest scores 82.9% on the LOCOMO benchmark — the standard evaluation for long-term conversational memory (ACL 2024).

Evaluated with Claude Sonnet 4.5 as the answer agent and Haiku 4.5 as the judge, using the industry-standard LLM-as-a-Judge methodology. All 5 LOCOMO categories included.

Re-measured on 0.24.1 (same protocol, 199 questions): 84.4% and 85.4% across two runs of the default legacy fusion. Note the run-to-run noise — two runs of the identical configuration flipped 16 individual questions and differed by 1.0 point, so treat sub-2-point differences on this benchmark as inconclusive.

Fusion modes

MEMORY_FUSION=rrf (reciprocal rank fusion) exists because summing raw cosine with max-normalized BM25 adds incomparable scales. It fixes three measured scoring artifacts — see the 0.22.0 notes — but it did not improve answers, so legacy remains the default:

The retrieval-only numbers are deterministic and show the real trade: rrf surfaces considerably more gold evidence inside the top 20 (+8.1 points recall) but ranks it slightly lower (−0.008 MRR). Because the answer agent already reads the top 20, the extra recall didn't convert into better answers, and an independent A/B on a 38-fact corpus lost two answers outright to top-rank precision.

Two further costs, both measured:

• Scores stop discriminating. rrf inflates absolute scores (top-1 rises

0.70 → 0.93) and compresses their spread to ~0.01 across the top 4, versus ~0.17 under legacy, because rank 1 contributes 1.0 per channel however weak the match is. Exact ties between adjacent results are normal. In practice you cannot threshold on an rrf score, and ordering inside the band is decided by the tiebreak rather than by relevance.

• No absolute quality signal. Neither mode separates answerable from

unanswerable questions by score, so this isn't a lost refusal signal — but an rrf score carries no information about how good the match actually is.

• Superseded memories can be demoted out of the window. The supersession

penalty is a score multiplier (×0.5), which is proportionate against legacy's ~0.17 spread and heavy against rrf's ~0.01: under rrf a superseded memory usually falls below unrelated results rather than merely below its own correction. That is still a correct demotion — the current version ranks first in both modes — but it means rrf cannot show you a stale version for comparison. The pathological case, where the penalty hid the only relevant memory, applied to supersession cycles and is fixed: cycle members are exempt (see 0.28.2). A rank demotion rather than a multiplier would be the coherent analogue under rank fusion.

Use rrf when you want maximum recall in a window you will read entirely and stability under corpus edits. Keep legacy when you want scores that mean something, which is why it is the default.

Architecture advantages

• Zero LLM calls in the server — intelligence lives in the agent, not the memory layer
• Local embeddings — no API key needed (bge-small-en-v1.5, 384-dim)
• Single embedded database — no Docker, no PostgreSQL, no separate vector DB
• Hybrid search — Vector (HNSW) + Full-text (BM25) + Graph (PageRank + Louvain communities)
• Minimal retrieval surface — the benchmark agent above scored 82.9% using only memory_search, memory_get and a calculator for date arithmetic. Retrieval quality comes from the server, not from agent-side orchestration.

Quick Start

# Run directly with uvx (no install needed)
uvx memnest-mcp

Or install and run:

pip install memnest-mcp
memnest-mcp

Configure a project (Kiro)

From your project root, one command writes the workspace-level MCP config (with the memory scope pinned to the project) plus the recall/persist/dream agent hooks:

memnest-mcp config kiro            # configure the current directory
memnest-mcp config kiro --check    # verify only
memnest-mcp config kiro --no-hooks # MCP server config only

This writes .kiro/settings/mcp.json (server + pinned workspace), .kiro/hooks/memnest-recall.json and memnest-persist.json (automatic recall and persistence), and .kiro/steering/memnest-dream.md — a manual steering file you invoke with /memnest-dream to consolidate memory.

The config is always workspace-level (<project>/.kiro/), so each project gets its own correctly-scoped memory database at <project>/.memnest/. Reconnect MCP servers in Kiro afterwards. The Kiro Power (below) remains optional on top for keyword activation and skills.

MCP Configuration

Add to your MCP client config (Kiro, Claude Desktop, Cursor, etc.):

{
  "mcpServers": {
    "memnest": {
      "command": "uvx",
      "args": ["memnest-mcp@latest"],
      "env": {
        "FASTMCP_LOG_LEVEL": "ERROR"
      }
    }
  }
}

That's it — zero config required. All settings have sensible defaults.

Tools

memory_stats is also safe to paste by default: the workspace and database paths are reduced to basename#hash, which keeps identity (two calls on one database match, different databases differ) without naming your directory tree. Every derived diagnostic — db_inside_workspace, private_to_workspace, the workspace count — is computed server-side and unaffected. Pass include_paths=True when debugging locally.

Exports are safe to share by default: workspace values are absolute paths, so they are replaced with opaque labels (workspace-1) that preserve the distinction between workspaces without disclosing directory layout. Pass include_workspace_paths=True for a local backup where you want the real values. Import ignores the field either way and assigns the current workspace.

Graph Data Model

(:Memory)  — content, embedding, category, tags, importance, access_count, timestamps
(:Topic)   — auto-created from tags

(:Memory)-[:ABOUT]->(:Topic)          # memory is about a topic
(:Memory)-[:RELATED_TO]->(:Memory)    # memories are related
(:Memory)-[:SUPERSEDES]->(:Memory)    # newer memory replaces older

Example: Store and Search

# Store a memory (via MCP tool call)
memory_store(
    content="User prefers Python over Node.js for backend tools",
    category="preference",
    tags=["python", "nodejs", "backend"],
    importance=4
)

# Search memories
memory_search(query="what language does the user prefer")

# Traverse the graph
memory_query(
    cypher_query="MATCH (m:Memory)-[:ABOUT]->(t:Topic {name: 'python'}) RETURN m.content"
)

Example: Graph Relationships

# Link related memories
memory_relate(from_id=5, to_id=3, relationship="RELATED_TO")

# Mark a decision as superseded
memory_relate(from_id=8, to_id=2, relationship="SUPERSEDES")

# Find all memories about a topic
memory_query(
    cypher_query="MATCH (m:Memory)-[:ABOUT]->(t:Topic) RETURN t.name, COUNT(m) ORDER BY COUNT(m) DESC"
)

Three-Layer Deduplication

Every memory_store call runs through three dedup layers:

1. Exact hash — SHA256 of normalized content. Identical content is rejected, importance bumped.
2. Semantic similarity — If cosine similarity > 0.92 with an existing memory, merges into it (keeps longer content, merges tags, bumps importance).
3. Consolidation — Periodic via memory_dream. Auto-prunes stale low-importance memories, auto-merges trivial duplicates (similarity ≥ 0.95), surfaces clusters for LLM-driven review.

Categories

Configuration

All settings are optional — defaults work out of the box.

In-Memory Mode (Testing)

"env": { "MEMORY_DB_PATH": ":memory:" }

All data is ephemeral — lost on restart. Useful for testing.

Kiro Power

This repo includes a ready-to-use Kiro Power in the power/memnest/ directory, packaged in the Agent Plugins v1.0.0 format with:

• Plugin manifest with activation keywords (power/memnest/plugin.json)
• Pre-configured MCP server (power/memnest/mcp.json)
• Two Kiro agent hooks for automatic recall and persistence (power/memnest/dev.kiro/hooks/, v1 hook schema — IDE/CLI only; on Kiro Web the agent follows the same workflow from the getting-started skill)
• memnest-recall (UserPromptSubmit) — searches memory before responding to each prompt
• memnest-persist (Stop) — stores important info when the agent finishes
• consolidation runs on demand via memory_dream (see the getting-started skill)
• Agent Skills with the setup guide and Cypher query examples (power/memnest/skills/)

Install in Kiro: Add Custom Power → https://github.com/arunkumars-mf/memnest-mcp/tree/main/power/memnest

Architecture

AI Agent (Kiro, Claude, etc.)
    │
    ├─ memory_store ──→ embed content → dedup check → insert node → link topics
    ├─ memory_search ─→ embed query → HNSW vector search → tag boost → rank
    ├─ memory_query ──→ execute Cypher → return graph results
    │
    └─ LadybugDB (embedded, single directory)
        ├─ Memory nodes (content + FLOAT[384] embeddings)
        ├─ Topic nodes (auto-linked from tags)
        ├─ HNSW vector index (cosine similarity)
        └─ Graph relationships (ABOUT, RELATED_TO, SUPERSEDES, EXPLAINS)

Requirements

• Python 3.10+
• Dependencies installed automatically: real-ladybug, fastembed, mcp
• ~130MB disk for the embedding model (downloaded on first run)

TOON Format (Optional)

Memnest supports TOON (Token-Oriented Object Notation) as a response format, reducing token usage by 30–60% compared to JSON. This is useful when memory results are fed back into LLM context.

TOON is optional — the server falls back to compact JSON automatically if the package isn't installed. To enable it:

pip install "memnest-mcp[toon]"

Or with uvx (requires the --prerelease=allow flag since toon-format is currently in beta):

uvx --prerelease=allow --with "toon-format==0.9.0b1" memnest-mcp@latest

To switch formats at runtime, set the environment variable:

MEMORY_RESPONSE_FORMAT=toon   # compact, token-efficient (default when installed)
MEMORY_RESPONSE_FORMAT=json   # standard JSON (default when toon is not installed)

The official Python implementation of TOON is toon-format/toon-python, currently at v0.9.0-beta.1. Once it reaches a stable 1.0 release, the --prerelease=allow flag will no longer be necessary.

Contributing

Issues and PRs welcome. See LICENSE for terms.

License

MIT

Changelog

0.32.0

Batch shapes audited across all 20 tools. Five already took batches — memory_store (items), memory_update (updates), memory_relate and memory_unrelate (relations), memory_delete (memory_id as int or list). Two gaps remained, and they were not equivalent.

• `memory_keep_separate` accepts `pairs=[[a, b], [c, d]]`. It already took a list, which made it look batched, but the list is a clique: every pair among the ids is recorded. So an agent resolving three memory_dream clusters in one call by passing all their ids would record verdicts on cross-cluster pairs it never examined — permanently suppressing genuine conflicts between them. That is a silent correctness loss dressed as an optimisation, which is worse than having no batch path at all. memory_ids keeps clique semantics for a single cluster (the right shape when all members are mutually distinct); pairs records exactly what it is given. Both may be passed, the union is deduplicated, and [b, a] is recognised as the same verdict as [a, b].
• `memory_get` accepts a list, matching memory_delete's int-or-list parameter. Every surface that reports ids — conflict flags, review clusters, supersession cycles — reports several at once, so reading them back was N round trips. A single int returns the memory object unchanged; a list returns {"results": [...], "count": n, "not_found": [...]}, so existing callers see no difference. The batch path shares one _get_one helper with the single path rather than duplicating it: a second copy is how the access_count drift that _recreate_memory_node exists to end began, and a batch path that quietly lost the superseded flag would present a stale memory as current.

Deliberately not batched: memory_search (batching queries raises result-merging questions a caller can answer better than the server) and memory_export/memory_import (one file per call is the unit).

The dream steering, the persist hook and the getting-started skill now name pairs where they resolve more than one cluster, since that is the workflow in which the clique shape is actively wrong.

0.31.3

The hooks, the skill and the auto-approve lists were still teaching the pre-0.30.0 option set. 0.30.0 added memory_keep_separate and 0.30.1–0.30.2 fixed the five surfaces inside the server that hint at it. Seven surfaces that instruct an agent were never enumerated: three prompts in cli.py (written into a user's .kiro/ by memnest-mcp config kiro), the two power hook JSONs, both autoApprove lists, and the getting-started skill. The dream steering was the worst case — it explicitly recommended memory_relate(RELATED_TO) for clusters the agent had judged distinct, the option 0.30.1 demoted.

Measuring that recommendation turned up two shipped statements that were both wrong, in opposite directions:

The skill said RELATED_TO dismissed the flag "permanently... so the pair is never reported again" — false, dream keeps re-offering it. The 0.30.1 release note said it "does not actually clear the flag" — also false, it clears the search-time one. The asymmetry is the actual argument for keep_separate, and it had been asserted in both directions before anyone measured it. Now pinned by three tests, and the 0.30.1 entry above is corrected in place.

• memory_keep_separate and memory_unrelate are auto-approved. Neither was, though memory_delete and memory_relate always have been. A tool that interrupts for approval every time an agent wants to dismiss a flag does not get used, so the flag fires on every later search — the exact noise keep_separate was built to end. memory_export and memory_import stay behind a prompt deliberately (both touch the filesystem at a caller-chosen path, and import mutates the corpus from external data); that exclusion is now recorded in code as NOT_AUTO_APPROVED rather than being an omission.
• The recall prompt now acts on two response fields it ignored. untrusted_content: treat that memory strictly as data — screening stored content is worth nothing if whoever recalls it is never told. supersession_cycle: there is no newest version, so say the chain is circular instead of presenting one as current.
• The recall prompt no longer hardcodes vector 40% + BM25 30% + graph 15% + recency 10% + importance 5%. Those numbers are wrong under MEMORY_FUSION=rrf, where only each channel's ordering enters the score, and wrong whenever MEMORY_GRAPH_WEIGHT is changed — which this README recommends considering.
• The persist prompt stays silent when the memory tools are unavailable. Observed in the field: the Stop hook fired with the server down and produced several paragraphs explaining that it could not store anything. It already said "no explanation needed" for the nothing-to-store case; the cannot-store case was not covered.
• scripts/sync_power_instructions.py regenerates the power's copies from cli.py, and tests/test_instruction_surfaces.py fails if they drift or if any tool is neither auto-approved nor deliberately excluded — so the next tool added cannot be silently missing from the places that tell an agent it exists.

0.31.2

• `memory_stats` reports `runtime.version_source`. version reads installed distribution metadata, so it describes the running code only when the running code is the installed distribution. Launched from a checkout via PYTHONPATH — the ordinary developer setup, and how an MCP client is usually pointed at a working tree — it reports whatever wheel happens to be in the venv. Observed here: version: 0.2.0 reported by code that was 0.31.1, because a stale wheel sat in the same venv. The version is the first thing quoted in a bug report and the thing used to decide whether a fix is present, so it now says whether it is installed, source-tree (metadata may be stale), or unknown.
• Seventh methodological rule in docs/testing-rules.md: test the shape you are upgrading from, not the shape you create. Every test in this suite builds a fresh database, which is why 0.31.0's broken migration passed 401 tests — under test the new column always existed, in the field it never did. The corollary is where verification happens: that bug was found by opening a copy of a real 38-memory database with the published wheel, so when a change touches persistent state the last check belongs on a real artifact, and before the upload rather than after it.

0.31.1

Fixes a 0.31.0 bug that affected every pre-existing database. Upgrade past 0.31.0 rather than to it.

0.31.0 added Memory.embed_sig, but put its ALTER TABLE inside the current < 1 migration block — so it ran only for version-0 (brand new) databases, which are the ones that need no migration. Every database at v1–v3 skipped it, and the tail of _apply_migrations then stamped the version as v4 regardless, recording a migration that never ran. Because those databases report a current version, no version-gated repair could ever reach them again.

The damage was silent in both directions, which is why 401 passing tests did not catch it. Queries naming the missing column raise a Binder exception; _count_stale_embed_sig swallowed it and returned None; healthy evaluated not (None or 0) as True; and _backfill_embeddings returned a zero report. So the new feature did nothing at all on precisely the databases it was written for, while memory_stats reported healthy: true next to stale_signature: null. Verified against a real 38-memory database: before the fix, stale_signature: None and scanned: 0; after, 38 stale rows found and re-embedded.

• The column is now ensured by probe (_ensure_embed_sig_column), before the version gate can short-circuit, so databases already mis-stamped v4 by 0.31.0 self-repair on open. Same reasoning as the post-delete census: check the thing, do not infer it from a counter that claims it happened.
• healthy now requires stale_signature == 0 rather than not (count or 0), so an uncomputable count no longer reads as green. That sentinel-as-ok pattern is one this project has now fixed four times, and 0.31.0 shipped a fifth instance directly beneath a comment claiming to prevent it.
• Three tests cover the migration path, which previously had none: every test builds a fresh database, so the column was always present under test and always absent in the field.

0.31.0

Three mechanisms borrowed from kirocrew's memory implementation. Each closes a silent failure — the kind this project already spends effort detecting — rather than adding a feature.

• Prompt-injection screening on stored and returned content. Memory content is whatever an agent decided to keep: a scraped page, a PR body, a log line. It is then replayed into another agent's context, in a later session, with no human in the loop — so a memory containing ignore previous instructions is a persistent, cross-session injection carrier, and the only "injection" this server screened for was Cypher. Flagged, never rewritten and never refused: the content may legitimately be about prompt injection (this project stores several such memories), and silently mutating a caller's data would be worse than the risk. memory_store returns untrusted_content with the matched markers while the caller still has context; memory_search screens what it is actually returning, which also covers memories written before screening existed.
• Per-row embedding signature. _verify_embedding_dim catches a dimension change, which is the loud case. A same-dimension model swap is silent: vectors stay present, reachable, and individually valid while being mutually meaningless — similarity degrades and every health field still reads green, because the index is intact and only the vectors' meaning changed. Each row now records the model that embedded it (memory_stats → runtime.embeddings.signature / stale_signature), making drift detectable and repairable per row instead of requiring a full rebuild. healthy folds the count in, so it cannot read true beside stale_signature: 2.
• Embedding backfill (memory_dream, bounded at 25 rows/run, reported as embedding_backfill). This closes a hole with no prior repair path at all: when the model fails to load, memory_store creates the row without an embedding — memory_stats counted those rows and memory_reindex rebuilt the index, but nothing ever re-embedded the rows, so a memory stored during a model-load blip stayed invisible to vector search permanently. The per-query census could not flag it either, because it compares hits against rows that have embeddings. Backfill also repairs stale-signature rows. Bounded because each row costs an embed plus a delete+recreate (the engine cannot update an indexed embedding in place), and that churn is what damages the HNSW graph — so the post-delete census runs afterwards.

Schema v4 (Memory.embed_sig). Existing databases migrate on open and backfill lazily as memory_dream runs; no manual step. Not adopted from kirocrew, and why: soft deletes (the index keeps growing and every query must filter; the post-delete census already covers the failure they address), stale-episodic retirement (a destructive edit to stored history at a 0.7 cosine threshold — our worst bug class), and MMR diversity reranking (a real gap, but it changes ranking, so it belongs in its own measured pass).

0.30.2

• A fifth hint surface named the new tool. The write-time conflict hint has two variants — value_disagreement and near_duplicate — and 0.30.1's count of four treated them as one, because it was taken from the release description rather than from the emission sites in the code. The near_duplicate variant offered supersession or nothing, so an agent holding two facts that both hold could only wrongly mark one stale or leave the flag firing on every later search.

0.30.1

• Four hint surfaces named memory_relate(RELATED_TO) as the way to dismiss a conflict flag, which creates graph structure the agent did not intend. They now name memory_keep_separate, state that it records the verdict without creating an edge, and demote RELATED_TO to a conditional for genuinely connected memories. (Corrected in 0.31.3: this entry originally claimed `RELATED_TO` "does not actually clear the flag", which is wrong — see 0.31.3 for what it does.)

0.30.0

• `memory_keep_separate` makes an agent's leave_separate verdict durable. memory_dream offered that resolution and recorded nothing, so the judgement evaporated: the same cluster was re-offered next run and the same pair re-flagged on the next search. 0.25.0 reduced that noise by inferring permanent verdicts from the merge gates; this records the verdict an agent actually made, which is the case the gates cannot infer — two facts that read alike, share a subject, and genuinely both hold. Creates no edge (the distinction from memory_relate), is idempotent, and verdicts are collected automatically once either memory is deleted.

0.29.3

• Filesystem paths are scrubbed at the single serialization boundary, covering the ~17 sites that return engine error text verbatim. memory_export also returns filename, since the scrub reaches its own returned path.

0.29.2

• memory_stats redacts paths by default as basename#hash, keeping identity while every derived diagnostic (db_inside_workspace, private_to_workspace) stays valid. include_paths=True restores them.

0.29.1

• Exports no longer embed the workspace path (39 occurrences on a 38-memory export). Distinct workspaces become opaque labels, preserving multi-workspace fidelity. include_workspace_paths=True opts back in.

0.28.3

• `explain` no longer advertises a penalty it did not apply. Cycle members are exempt from the ×0.5 supersession multiplier, but the block still printed superseded_penalty: 0.5 next to a score that was never halved. It now reports null with superseded_penalty_exempt: "supersession_cycle", and a penalised row's score is asserted to equal sum(weighted) × penalty. superseded: true still shows, because each member genuinely is superseded — what changed is only whether a penalty was charged.

0.28.2

• Supersession-cycle members are exempt from the ×0.5 penalty. In a cycle every member is superseded by construction, so the flag says nothing about which is stale while the multiplier still destroys ranking. Measured: for one query the answering memory scored an unpenalised 0.769 (vector 0.857, FTS 1.0) and was halved to 0.385, below three unrelated memories at ~0.41 — so at top_k=2 it was not returned, and because the cycle warning is scoped to returned rows the warning vanished with it. The caller asking exactly the affected question got unrelated results and no indication anything was wrong. Cycle members now rank on relevance with supersession_cycle attached, which fixes the same failure under rrf, where the compressed score spread made it unavoidable.
• Ordinary correction chains are unaffected: stale versions are still demoted.

0.28.1

• `supersession_cycle` is scoped to the returned rows. It was keyed off the superseded set derived from scored candidates (the pool of 100), which on any workspace smaller than the pool is the whole corpus — so one unresolved cycle attached the warning to every unrelated search. Now it triggers only when a returned row is a cycle member, then reports the full cycle so the loop is repairable.
• Records two findings from the graph-channel investigation: PageRank does propagate (a hub with 7 incoming edges reached 4× the teleport floor) but only RELATED_TO feeds a memory incoming rank — ABOUT points Memory→Topic, and SUPERSEDES/EXPLAINS are excluded from the projection entirely, so on a corpus with few asserted RELATED_TO edges nearly every memory sits at the floor and k_degree is the only varying term. And under rrf the multiplicative supersession penalty makes superseded memories unreachable.

0.28.0

• `memory_dream(dry_run=True)` now reports contradictions. SCC detection on the SUPERSEDES subgraph was gated behind the write path, so a genuine cycle reported contradictions: [] on inspection and only surfaced when run for effect — a diagnostic reading clean on a state that isn't. Detection is read-only; there was never a reason for the gate.
• `memory_search` surfaces a supersession cycle. In a cycle no memory is current, so every member is equally penalised, the oldest value can rank first, and the documented current-answer query returns zero rows — which reads as "no information" rather than "contradictory information". A supersession_cycle field now names the members and the repair, alongside potential_conflicts.
• `MEMORY_GRAPH_WEIGHT` is configurable (default unchanged at 0.15). The graph channel is relevance-independent and dormant until dream runs; on a tag-dense corpus disabling it measured strictly better (hit@5 0.469 → 0.485, MRR@20 0.354 → 0.370). Deleting all 858 auto-inferred edges changed nothing, locating the effect in ABOUT edges to Topic nodes — i.e. tag popularity, not knowledge structure.
• `benchmark/anchors.py` pins retrieval anchors for both graph states. The long-used "anchors bit-identical" check was only valid on a corpus where dream had never run, so it would read as a regression on any real workspace. Cold and warm values are now recorded separately and both asserted.

0.27.0

• Result ordering no longer depends on the order memories were stored. Two fixes to the same defect class: the rrf rank transform broke channel-value ties by memory id (and ids encode insertion order, so identical BM25 scores produced arbitrary ranks that propagated into different fused scores), and the final sort broke score ties by dict order. Channel ranks now use competition ranking — equal values get equal rank — and final ties break on importance, then recency, then id. A 4-memory fixture that reordered its own results purely by store order now doesn't.
• This was costing rrf measurable quality: gold-evidence recall@20 rises 64.3% → 66.3% and recall@5 45.4% → 46.9% (now equal to legacy). legacy is unaffected — it does no rank transform, and anchors are bit-identical.
• Documents the rrf score-compression cost: spread across the top 4 is ~0.01 versus ~0.17 under legacy, so rrf scores cannot be thresholded.

0.26.2

• Verifies a real restore into a separate on-disk database, not just an in-memory one: embeddings recomputed, index fully reachable, restored memories findable, supersession still resolving to current.

0.26.0

Closes two silent coverage losses that appeared at ordinary corpus sizes, not extreme ones.

• Index-health coverage no longer lapses above the candidate pool. The per-query census compared ranked vector hits against the corpus, which only works while the pool (100) covers it — so the detector for the worst bug class protected a shrinking slice as a workspace grew (2% at 5,000 memories). Above the pool it now runs a dedicated id-only probe at k=corpus: measured 19.7 ms against a 191.8 ms search, ~10% overhead, 100% coverage at every size. explain_meta.census_mode reports which path ran.
• Dream's scan cap is a rotating window, not a horizon. It always examined the newest MEMORY_CONSOLIDATE_SCAN by updated_at, so once a workspace passed the cap everything older was never considered for merge or prune again. The window now advances each run and wraps, covering any corpus over successive runs at unchanged per-run cost. memory_dream reports scan_coverage.
• `memory_stats` reports `db_scope` — whether the one-database-per-workspace invariant that makes lock-free operation safe actually holds, judged from which workspaces own memories in the file rather than from configuration. Warns when a shared MEMORY_DB_PATH has put unrelated projects in one graph.
• Soak harness gains --sessions N: N short sessions as separate processes, each verifying what the previous one left behind. This is the realistic stress pattern for one connection per workspace, and it is where the delete-churn bug actually manifested.

0.25.0

• Dream review clusters now carry only pairs that still need a decision. Different-subject and disjoint-scope pairs are separate permanently and are dropped; same-subject value conflicts are unresolved, so they stay and are labelled gate: "value_conflict" with the non-destructive resolution named.

0.24.1

• memory_delete reaps orphaned Topic nodes (a long-lived database had accumulated 273 orphans against 24 live topics); memory_dream does the same for its own prune/merge deletions and reports topics_reaped.
• The post-delete index census is exception-isolated — diagnostics can never fail a committed write.

0.24.0

Root-caused the recurring silent loss of vector-search coverage: LadybugDB 0.15.3 delete maintenance progressively orphans surviving HNSW nodes when transient batches are inserted then deleted. Reproduced standalone (repro), monotonic, in-process, persists across restarts. Onset is non-monotonic in burst size and seed-dependent, which is why several earlier experiments wrongly cleared it.

• memory_delete now censuses index reachability and rebuilds on shortfall, so a session's deletes can't hand the next session a degraded index.
• memory_stats reports vector_index.status: "degraded" when the census contradicts the cached probe verdict.

0.23.0

• One shared implementation of the delete+recreate path used by store-dedup, update and dream merge (three copies had drifted; dream was zeroing access_count). Store and update preserve the count; dream merge sums the merged members'.

0.22.0

• New MEMORY_FUSION=rrf (opt-in). Fixes three artifacts of summing incomparable channel scales: survivor keyword scores rescaling 2.24× when the top hit was deleted, a +0.166 score jump that inverted a ranking, and a plateau of identical scores when the vector channel was dead. See Fusion modes for why it is not the default.

0.19.0

Surface hardening from a full tool-by-tool review.

• Security: the memory_query destructive-query guard was bypassable. It matched the substring "DELETE " — with a literal trailing space — so MATCH (m:Memory)\nDETACH\nDELETE\nm; reported success and deleted every memory with MEMORY_ALLOW_DESTRUCTIVE=false. Queries are now classified after stripping comments and string literals, matching keywords on word boundaries.
• Breaking: read_only=True now rejects any mutation. It previously permitted CREATE/MERGE/SET, so an overwrite succeeded under a flag named read-only. MEMORY_ALLOW_DESTRUCTIVE now also covers SET, REMOVE and COPY — an overwrite destroys the previous value as surely as a delete.
• New `memory_unrelate`: edges could be created but never removed, and memory_query's DELETE is blocked by default, so a mistaken SUPERSEDES was permanent. This is also the supported way to break a circular SUPERSEDES chain that memory_dream reports.
• `memory_get` now returns edges (include_edges=True by default) plus superseded / superseded_by. Answering "what does this replace?" no longer requires Cypher.
• Ranking fix: the search candidate pool was top_k * 3, so the page size decided which memories were scored at all — on a 25-memory corpus, top_k=10 surfaced two memories that outranked every result top_k=5 returned. The pool is now fixed (MEMORY_SEARCH_CANDIDATES, default 100) and independent of top_k. Scores are unchanged; only coverage improves.
• Pagination: memory_search(offset=...) with offset / has_more in the response. Rank 11+ was previously unreachable.
• New `memory_export` / `memory_import`: JSON backup and restore including edges, with id remapping so an import can merge into an existing database. memory_relate is now idempotent (status: "exists"), so re-importing no longer doubles every edge.
• Input validation across every tool: two-sided clamping (preview_chars=-5 used to slice content from the wrong end; top_k=0 returned a degraded flag blaming the embedding model), content and batch size caps, and honest statuses (memory_delete reported deleted when every id was missing; memory_list(min_importance='high') raised a raw ValueError).
• memory_get and memory_list are no longer labelled compatibility aliases — each does something no other tool does. memory_traverse is marked deprecated.

0.3.0

• Default database is now per-workspace: .memnest/memory.lbug in the current directory. No more cross-workspace lock conflicts.
• Set MEMORY_DB_PATH to use a custom location (e.g. ~/.memnest/memory.lbug for global shared memory).
• Hybrid search: Vector (HNSW) + Full-text (BM25) + Graph scoring with PageRank, Louvain community detection, and K-Core decomposition.
• LOCOMO benchmark: 82.9% overall score.

0.2.0

Compatibility-preserving redesign with improved safety defaults.

• New tools: memory_query (general Cypher), memory_schema, memory_topics, memory_dream, memory_graph_html. Batch mode added to memory_store, memory_update, memory_relate, memory_delete.
• Breaking: MEMORY_ALLOW_DESTRUCTIVE now defaults to false. Set it to true if you previously relied on memory_query deleting nodes.
• Breaking: tag storage migrated from comma-joined strings to JSON arrays. Old rows are still readable; rewriting (e.g. via memory_update) upgrades them to JSON.
• memory_get, memory_list, memory_traverse from 0.1.x are retained as compatibility aliases. (As of 0.19.0 memory_get and memory_list are first-class again; only memory_traverse remains deprecated.)
• TOON serialization is now the default response format when toon-format is installed; set MEMORY_RESPONSE_FORMAT=json to opt out.
• memory_relate validates that both endpoints exist before returning created (used to silently no-op on typo'd IDs).
• memory_graph_html is now XSS-safe (HTML-escaped tooltips, DOM textContent for the detail panel), refuses to render >MEMORY_GRAPH_MAX_NODES, and rotates snapshots.
• Workspace filter pushed inside the vector index WITH clause so search recall isn't starved across workspaces.
• Dream consolidation: dedupes parallel edges across merges, isolates clusters by workspace, persists state via atomic sidecar JSON.