keymem

The associative memory layer for LLM agents — recall by association, not just similarity.

keymem demo — asking about the party cake surfaces Mina's peanut allergy via a shared key

Most agent memory is a vector store. It surfaces what sounds like your query — and misses everything your query is connected to.

keymem stores memories in a key graph instead. A search for "Newton" can still reach "strawberries" — Newton → apple → fruit → strawberry. The path lives in the graph, not in embedding space. It runs locally as an MCP server, so any MCP-compatible agent gets human-like associative recall with no external database.

Works with: Claude Desktop · Claude Code · any MCP-compatible LLM agent

Why associative memory?

Vector-store memory retrieves by embedding similarity. That works until the thing you need isn't similar to the words you typed:

Query: "Newton"
Similarity search finds: "Newton discovered gravity"  ✅
Similarity search misses: "user likes strawberries"   ❌

A person makes the leap anyway — Newton reminds them of the apple, apples are fruit, they like strawberries. keymem makes that same leap because the path exists in the key graph: Newton → apple memory → fruit key → strawberry memory. No embedding distance connects "Newton" and "strawberry"; a chain of shared keys does.

This is the core idea: memories are not islands ranked by distance. They are nodes in an N:M key/value graph that an agent can walk.

How it works

Key Space (concepts)         Value Space (memories)
[apple]   ────────┬─────────→      ↑ same memory
[gravity] ────────┘
                  │
[apple]   ────────┼─────────→ "apples are red fruit"
[fruit]   ──────┬─┘
[red]     ──────┤
                │
[fruit]   ──────┼─────────→ "user likes strawberries"
[strawberry]────┘

Memories live in a Value Space, reached through a separate Key Space — one memory reachable via many keys, one key leading to many memories.

recall("Newton") returns matching key clusters such as [Newton] and [apple], not memory content. The agent then navigates explicitly: read_key(apple) → select the Newton memory → read_memory(...) → discover its [fruit] key → read_key(fruit) → select the strawberry memory.

The default MCP flow is therefore Key → Memory → Key. Full memory content enters the model context only when the agent deliberately calls read_memory() — so broad concepts never flood the context window.

Quick Start

keymem is an MCP server (a CLI), not a library. Run it with npx -y keymem (recommended — always the latest) or install the command globally with npm i -g keymem. Do not add it to your app with npm i keymem as a dependency: it bundles openai, zod, and the MCP SDK, so inside an existing project it just duplicates those trees (and can clash with your app's zod/ openai versions). The npm i keymem line npm shows on the package page is for libraries — it doesn't apply here.

# Optional global install (npx needs none). This puts a `keymem` command on PATH that
# MCP clients can spawn. Run bare, it starts a stdio MCP server and waits for a client —
# so point your MCP config at `keymem` (or just use `npx -y keymem` as shown below).
npm i -g keymem

Claude Desktop

Add to claude_desktop_config.json:

OpenAI embeddings:

{
  "mcpServers": {
    "keymem": {
      "command": "npx",
      "args": ["-y", "keymem"],
      "env": {
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}

Local embeddings (no API key required) — bge-m3 recommended:

{
  "mcpServers": {
    "keymem": {
      "command": "npx",
      "args": ["-y", "keymem"],
      "env": {
        "EMBEDDING_BACKEND": "local",
        "LOCAL_EMBEDDING_MODEL": "bge-m3"
      }
    }
  }
}

bge-m3 (multilingual, recommended) auto-downloads ~570MB on first run, then caches. Omit LOCAL_EMBEDDING_MODEL for the lighter default (fast-multilingual-e5-large). Add "KEYMEM_RERANK": "true" to enable cross-encoder reranking (downloads a second model on first use).

Claude Code

# OpenAI embeddings
claude mcp add keymem -e OPENAI_API_KEY=your-key -- npx -y keymem

# Local embeddings (no API key required) — bge-m3 recommended (auto-downloads ~570MB on first run)
claude mcp add keymem -e EMBEDDING_BACKEND=local -e LOCAL_EMBEDDING_MODEL=bge-m3 -- npx -y keymem

That's it — recall and remember work immediately. The agent calls recall before its first reply, navigates with read_key/read_memory, and saves with remember. To wire in the recommended behavior (recall silently, use diverse keys, never mention the memory system to the user), include the memory_system_prompt MCP prompt in your system prompt.

Manual / Development

git clone https://github.com/donggyun112/keymem
cd keymem
pnpm install

Create .env:

OPENAI_API_KEY=your-openai-api-key
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

Or use local embeddings (no API key required):

EMBEDDING_BACKEND=local
LOCAL_EMBEDDING_MODEL=fast-multilingual-e5-large  # default; best fit for Korean/multilingual keys

pnpm dev
# or:
pnpm build
pnpm start

Requirements:

• Node.js 20+
• pnpm for local development
• OpenAI API key for OpenAI embeddings, or fastembed for local embeddings

Features

• N:M key/value graph — memories and the concepts that index them are separate spaces, linked many-to-many. One memory is reachable through many keys; one key leads to many memories.
• Agent-driven Key → Memory → Key navigation — the agent walks the graph deliberately instead of collapsing it into one opaque similarity search.
• Associative multi-hop recall — reach memories no embedding distance would connect, by following chains of shared keys.
• Depth system — every memory has a stability score 0.0 → 1.0. Frequently recalled facts deepen, stabilize, and decay slower.
• Versioning, not overwriting — corrections preserve the full history (when a belief changed, and from what).
• Key types — concept keys match by similarity; name/proper_noun keys match exactly, so "동건" never matches "뉴턴" just for being short.
• Cross-lingual key merging (IDF) — 파이썬 and Python collapse into one canonical cluster instead of fragmenting the key space.
• Hebbian link learning — the path an agent actually traverses gets reinforced ("fire together, wire together"), so useful associations become easier to reach.
• Hybrid retrieval (optional direct mode) — BM25 + dense + Reciprocal Rank Fusion, with depth/time modulation and configurable multi-hop expansion.
• Cross-encoder reranking (opt-in) — bge-reranker-v2-m3 re-scores candidates in direct mode.
• Local-first — all data in a local JSON graph; no external database. OpenAI or fully-local embeddings (auto-downloaded).

Depth System

Every memory has a depth score 0.0 → 1.0:

Depth increases +0.05 each recall. Deep memories decay slower over time. If you try to correct a deep memory, it resists — its depth stays higher even after supersede.

Key Types

Not all keys should behave the same. Names shouldn't match semantically — "동건" shouldn't match "뉴턴" just because they're both short Korean words.

Name/proper_noun keys also get an IDF penalty (×0.5) when they become hub keys connected to many memories, preventing them from polluting unrelated searches.

Versioning (not overwriting)

"user lives in Seoul"   (depth: 0.4 → weakened to 0.12, preserved)
        ↑ superseded by
"user moved to Busan"   (depth: 0.0, new)

keymem keeps the full history instead of overwriting on change. Every correction is traceable — when did the belief change, and from what session?

Key Merging

Add key "파이썬"  → finds existing "Python" (similarity 0.87 > threshold 0.85)
                 → reuses existing key instead of creating duplicate

Prevents key space fragmentation. The same concept across languages or phrasing stays unified.

Agent-driven Retrieval (default)

The default MCP API keeps Key Space and Value Space separate:

1. recall(query) searches canonical keys and aliases. It returns key IDs, concept labels, match scores, linked-memory counts, hub status, and specificity — never memory content.
2. read_key(key_id) returns ranked memory IDs and metadata, never content. Hub keys are paginated with limit/offset so broad concepts cannot flood context.
3. read_memory(memory_id, via_key_id) returns the full memory plus every connected key cluster. Only this full read increases memory depth/access count; only the traversed via_key_id edge receives Hebbian reinforcement.
4. The agent follows any returned key with another read_key() call, producing an explicit Key → Memory → Key graph walk.

Semantically merged keys are preserved as aliases on one canonical key cluster (for example Python + 파이썬). A key linked to at least three active memories is surfaced as a hub with is_hub, memory_count, and specificity metadata rather than being hidden by IDF. Override the threshold with KEYMEM_KEY_HUB_MIN_LINKS.

Direct Hybrid Retrieval (optional compatibility mode)

Set KEYMEM_DIRECT_RECALL=true to expose recall_memories(), a one-call memory retrieval path. Three signals run in parallel and are fused with Reciprocal Rank Fusion (RRF_K = 60):

• BM25 (sparse): lexical full-text search over memory content (MiniSearch, fuzzy + prefix). Catches exact terms, names, and rare tokens that embeddings blur.
• Dense Path A (key matching): query embedding → match keys → follow links → memories. Score = keySim × IDF × linkWeight, summed across all matching keys.
• Dense Path B (content matching): query embedding → directly compare against memory content embeddings. Finds memories even when they weren't tagged with the right keys.

Sparse and dense rank lists are merged by RRF, then modulated by depth and time before configurable multi-hop expansion (hops=1–5, default 2). This compatibility tool is hidden by default so agents use explicit key navigation instead of collapsing the graph into one search call.

Hebbian Link Learning

Reading a full memory is a write, not just a read. In the default flow, recall() and read_key() are read-only; read_memory(memory_id, via_key_id) reshapes the selected path:

• The traversed via_key_id → memory_id link is reinforced (+0.1, capped at 3.0).
• Memory depth and access count increase only after the agent reads the full memory.

Reinforcement is scoped to the key the agent actually traversed — not every key attached to the memory. This is the literal Hebbian rule ("fire together, wire together") and prevents unrelated associations from growing when the memory is reached through a different concept. Weights are clamped to [0.1, 3.0].

Link weights feed back into read_key() ranking, so repeatedly selected paths become easier to reach. Optional recall_memories() retains the previous matched-link reinforcement and explored-link decay behavior.

Architecture

│                      Key Space                          │
│   [name] [동건] [programming] [python] [fruit] [red]   │
│      ↓      ↓         ↓           ↓       ↓      ↓     │
│   [vec]  [exact]    [vec]       [vec]   [vec]  [vec]   │
                         │ N:M links
                         ↓
│                     Value Space                         │
│   "user's name is Donggeon"     depth: 0.85  (deep)    │
│   "user likes Python"           depth: 0.30  (medium)  │
│   "user likes strawberries"     depth: 0.05  (shallow) │

Default MCP navigation:

1. Embed the query and match canonical key concepts plus exact aliases.
2. Return key clusters with memory_count, is_hub, and specificity; do not return memory content.
3. Rank a selected key's memory handles by link weight, depth, and time decay in read_key().
4. Return full content and adjacent key clusters from read_memory(); reinforce only the traversed edge.
5. Repeat read_key(next_key_id) to walk the graph deliberately.

Optional `recall_memories()` algorithm (hybrid, configurable 1–5 hops; default 2):

Three retrieval signals run in parallel, then get fused and expanded:

1. BM25 (sparse): lexical search over memory content (MiniSearch, fuzzy 0.2 + prefix). Keep top 50.
2. Dense Path A (keys): embed query → match keys (concept: cosine ≥ threshold; name/proper_noun: substring match → score 1.0) → take top 10 keys → follow links. Score = keySim × IDF × linkWeight, summed across matching keys.
3. Dense Path B (content): compare query embedding directly against memory content embeddings (cosine ≥ threshold).
4. RRF fusion: merge the BM25 and dense rank lists via score += 1 / (RRF_K + rank + 1) (RRF_K = 60).
5. Depth & time modulation: score × (0.9 + depth × 0.1) × timeFactor, where timeFactor is a depth-weighted 30-day half-life decay (deep memories decay slower).
6. Associative expansion (`hops`, default 2): breadth-first from the directly-matched set — each round follows shared keys (× HOP_DECAY(0.3) × IDF × linkWeight) and explicit related_to links (bidirectional, × HOP_DECAY) to the next frontier. hops=N walks up to N steps, so a memory's hop is its shortest chain distance. Score decays by HOP_DECAY per hop.
7. Hebbian update: reinforce matched-key links of returned memories (+0.1), decay explored-but-unreturned links (−0.005).
8. Return ranked results with hop field (1 = direct, 2+ = associative distance).

Similarity thresholds (calibrated per embedding model)

Embedding backends have very different cosine distributions, so a single threshold set cannot serve all of them. The thresholds below are calibrated per model (getThresholdProfile() in src/embedding.ts):

Why e5 differs so much: multilingual-e5 packs embeddings into a narrow high-cosine band (~0.86–0.99). Same-word query↔key pairs (asymmetric query:/passage: prefixes) still separate cleanly (~0.89 vs ≤0.82), but key↔key and content↔content do not — distinct facts like "A uses Postgres" and "B uses Mongo" sit at ~0.96, dangerously close to true paraphrases (~0.99). Hence e5's merge/dedup/auto-link thresholds are pushed high to avoid silently collapsing distinct memories.

Drift escape hatch: if you switch models or your data's character drifts, override any threshold without code changes:

KEYMEM_KEY_RECALL=0.82
KEYMEM_MEMORY_DEDUP=0.99
# also: _KEY_MERGE, _KEY_AUTOLINK, _CONTENT_RECALL  (values in [0,1])

Score gate, distribution gate, and contradiction band can also be tuned per deployment:

Why e5 gates are opt-in: multilingual-e5's narrow cosine band (~0.86–0.99) makes a static floor unreliable, while held-out tests showed distribution and key-proximity gates can also overfit. Both are disabled by default to avoid hiding real memories. Use bge-m3 for reliable not-found behavior, or calibrate e5 gates on your own corpus.

Distribution gate parameters:

• `gateZ` — set via KEYMEM_GATE_Z or the min_z parameter of optional recall_memories().
• `0` disables the gate (default for bge-m3, bge, openai, minilm — where min_score already works).
• Both gates compose (AND): a result must clear both min_score and gateZ to be returned.
• A literal name/proper-noun key match (e.g. querying a stored name-typed key exactly) is always a definite anchor and bypasses the distribution gate.
• `GATE_MIN_POPULATION = 8`: the gate is skipped when fewer than 8 memories exist (too few samples for a reliable distribution), so early-session recall is unaffected. The gate's background population is namespace-filtered and excludes superseded/expired memories, so recall scoped to a sparse namespace may fall below this threshold and skip the gate entirely.
• Known e5 limitation: the optional gate keys off maxContentSim (content cosine only). A relevant fuzzy-key hit with a flat content distribution may be rejected; literal key matches bypass the gate.

An uncalibrated LOCAL_EMBEDDING_MODEL falls back to the BGE profile and logs a warning so the miscalibration is never silent.

Multilingual note: cross-lingual content matching has a same-language bias (a Korean query scores Korean memories higher regardless of meaning). The reliable cross-lingual path is the key graph — tag memories with keys in multiple languages (e.g. ["딸기", "strawberry"]) so recall hits the key exactly instead of relying on biased content similarity.

MCP Tools

The memory system exposes 12 tools by default:

Set KEYMEM_DIRECT_RECALL=true to expose a thirteenth compatibility tool, recall_memories(...), with BM25+dense+RRF multi-hop behavior.

A system prompt template is also available via the memory_system_prompt MCP prompt — include it to instruct the agent to recall silently, use diverse keys, and never mention the memory system to users.

Local embedding models

BGE-M3 (recommended, multilingual) — auto-downloaded: set LOCAL_EMBEDDING_MODEL=bge-m3 (aliases: bgem3, baai/bge-m3, fast-bge-m3). On first use the model is fetched automatically if missing — quantized ONNX (~570MB, from onnx-community/bge-m3-ONNX) plus the tokenizer/config (from BAAI/bge-m3) — and cached under ~/.keymem/models/bge-m3. No manual download needed:

EMBEDDING_BACKEND=local
LOCAL_EMBEDDING_MODEL=bge-m3
# optional — point at an existing model dir to skip the download (backward compatible):
# LOCAL_EMBEDDING_MODEL_PATH=/absolute/path/to/model-dir   # dir with model.onnx + tokenizer files
# LOCAL_EMBEDDING_MODEL_FILE=model.onnx                    # optional; default is model.onnx

First run downloads ~570MB once, then reuses the cache. If LOCAL_EMBEDDING_MODEL_PATH already holds the model it is used as-is with no download (a partial dir is self-healed — only missing files are fetched). Online-API backends (OpenAI) and fastembed built-ins are unaffected.

Cross-encoder reranking (optional direct mode): set KEYMEM_DIRECT_RECALL=true and KEYMEM_RERANK=true to re-score recall_memories() candidates with bge-reranker-v2-m3. The default key-navigation flow does not load the reranker. The model (~570MB, quantized) auto-downloads on first use and caches under ~/.keymem/models/reranker.

KEYMEM_RERANK=true
KEYMEM_DIRECT_RECALL=true
# optional: KEYMEM_RERANK_MODEL_PATH=/dir   KEYMEM_RERANK_POOL=30  (candidates re-scored)

Off by default. If the model cannot load, recall_memories() falls back to fused ranking. Query decomposition remains the caller's responsibility.

Reranker not-found gate (`KEYMEM_RERANK_MIN_SCORE`): in direct compatibility mode, reject the complete recall_memories() result when the top cross-encoder logit is below this floor.

KEYMEM_RERANK=true
KEYMEM_DIRECT_RECALL=true
KEYMEM_RERANK_MIN_SCORE=0   # reject when top rerank logit < 0 (bge-reranker-v2-m3 scale)

⚠️ Caveats. (1) Unset by default — no gate. (2) The logit scale is model-dependent; 0 (≈ sigmoid 0.5) suits bge-reranker-v2-m3 (measured: same-language found ≈ +3.9, not-found ≈ −5 to −6) — recalibrate for other rerankers. (3) Trusted for SAME-LANGUAGE only. Cross-lingual relevance logits run low even when relevant (KR query ↔ EN memory ≈ −5.4), so the gate auto-bypasses on a script mismatch (KR↔Latin) to avoid false-rejecting cross-lingual hits — which means cross-lingual content must be reachable via bilingual keys (["Jiwoo","지우"]), and cross-lingual not-found precision is a known limitation. Leave this off if you can't tag bilingual keys.

Prefix behavior: BGE-M3 does not use passage:/query: prefixes — embeddings are passed through as-is. All other local models (e5, BGE-en, MiniLM) continue to use prefixes unchanged.

Recommended for multilingual / cross-lingual use: `bge-m3`. It separates unrelated queries more reliably and performs substantially better than e5 on the project's Korean↔English fixtures. In optional direct mode, bge-m3's absolute min_score gate reaches ≈96% on the gate fixture; e5 requires corpus-specific tuning.

If OPENAI_API_KEY is not set and EMBEDDING_BACKEND is unset, the server automatically uses the local fastembed backend. For English-only use or lower local resource usage, set LOCAL_EMBEDDING_MODEL=fast-bge-base-en-v1.5 or fast-bge-small-en-v1.5.

Switching backends is safe. The graph records an embedding fingerprint (backend + model id) identifying the vector space its embeddings live in. On startup, if the current backend's fingerprint or dimension differs from what is stored, the graph auto-migrates — every key and memory is re-embedded with the new backend while content, links, depth, and access history are preserved (a graph.json.bak.* backup is written first). The fingerprint matters because two models can share a dimension yet produce incompatible vectors (e.g. fast-multilingual-e5-large and bge-m3 are both 1024-d); a dimension check alone would miss that swap and silently corrupt every similarity. Disable with KEYMEM_AUTO_MIGRATE=false. Re-embedding via OpenAI incurs one-time API cost proportional to your memory count.

>

Migrating a pre-fingerprint (legacy) graph across same-dimension models. A graph written before fingerprinting has no recorded vector space, so a same-dimension model swap off it cannot be detected automatically. Set KEYMEM_FORCE_REEMBED=true for one startup to re-embed unconditionally and stamp the fingerprint; remove it afterward (left on, it re-embeds on every start). This is exactly the one-shot needed when moving an existing e5 graph to bge-m3.

Data Storage

All data is local. No external database required.

~/.keymem/
├── graph.json          # canonical keys, aliases, memories, weighted links
└── conversations/
    └── {session_id}.jsonl   # original conversation turns

Set KEYMEM_DATA_DIR to use a different storage directory.

Benchmarks

On HotpotQA bridge questions (real external multi-hop data, gold labels, no LLM judge — and keys generated blind by independent subagents that never saw the question or answer), the key-graph retrieves both gold supporting paragraphs 63% of the time vs 53% for flat semantic retrieval and 35% for lexical (+10pp / +28pp) — the connected-but-dissimilar case it's built for. (Honest: my own hand-derived keys inflated this to 78/60; on non-multi-hop "comparison" questions the graph slightly hurts; it's retrieval-recall, not answer accuracy.) The read path is also O(1) (read_memory p50 ~45ms → ~0.01ms @ 500 memories). Full methodology and caveats: [BENCHMARKS.md](BENCHMARKS.md).

Limitations

• Linear scan — suitable for personal use (~10k memories). FAISS/ChromaDB integration planned for larger scale.
• Agentic round trips — the default recall → read_key → read_memory flow is more controllable and context-efficient, but needs more tool calls than one-shot retrieval.
• Hub breadth — broad keys can connect many memories. read_key() paginates hubs; the agent must choose whether to continue paging or follow a more specific adjacent key.
• Agent quality matters — key selection on remember affects retrieval quality. System prompt tuning is important.
• Cross-lingual content bias — with multilingual e5, raw content similarity favors same-language memories regardless of meaning. Tag memories with multilingual keys so the key graph (not biased content cosine) carries cross-lingual recall.
• Threshold calibration — thresholds are tuned per embedding model. A new/uncalibrated model falls back to the BGE profile (with a warning); recalibrate via the KEYMEM_* env overrides.

Testing

pnpm test                        # unit tests (fast, no model download)
tsx test/scenarios.ts            # 21 end-to-end behavioral checks (local e5)
tsx test/robustness.ts           # threshold overrides + Hebbian pollution bounds
tsx test/migration.ts            # backend/dimension switch auto-migration (no brick)
tsx test/nhop.ts                 # N-hop chained traversal (recall hops parameter)
tsx test/depth-noise.ts          # deep-hop noise bounds + relative score floor
tsx test/live-multilingual.ts    # interactive multilingual recall demo

# Manual retriever-quality check (NOT part of pnpm test):
EMBEDDING_BACKEND=local npx tsx test/retriever-quality.live.ts
# For bge-m3: also set LOCAL_EMBEDDING_MODEL=bge-m3 LOCAL_EMBEDDING_MODEL_PATH=/abs/dir

scenarios.ts and robustness.ts exercise the real local embedding backend (direct/associative/cross-lingual recall, versioning, depth growth, dedup, TTL, Hebbian learning, namespace isolation). They double as a recalibration harness when tuning thresholds for a new model.

Roadmap

• [ ] FAISS/ChromaDB for scale
• [ ] Coding agent profile (different key strategies for code context)
• [ ] Memory export/import
• [ ] Multi-user support

Author

donggyun112 — github.com/donggyun112

Repository: donggyun112/keymem · Issues & PRs welcome.

License

MIT © donggyun112