lore

lore is your long-term memory: a local SQLite + FTS5 store of agent session transcripts at ~/.lore/lore.db, shared by every harness on this machine (claude-code, codex, cursor, …). This skill drives the whole app from the CLI. The lore command opens the SQLite file directly, so every task here — find, read back, navigate, write — works with no server and no client. The MCP server (lore serve) is just one more reader of the same store; you never need it to use lore.

If lore help errors, the CLI isn't installed yet — and if a search comes back empty because a harness was never indexed, that's the onboarding job. Both are covered by this same skill: read `references/setup/index.md` to install lore, index or backfill a harness, teach lore a new format, or feed a live process, then come back here to use what you indexed. This skill self-bootstraps the low-level CLI substrate; for higher-level workflows, route to the workflow skills below instead of re-inventing their output contracts.

Workflow skills built on top of lore

Use this lore skill for the raw substrate: install, index, sync, search, get, context, session navigation, push, and destructive-memory safety. When the task is not just retrieval mechanics, hand the work to the workflow skill that owns the behavior:

• lore-recall (lore:recall in product docs): use for bounded memory answers, query plans, freshness labels, evidence packets, and retrieval-failure ledgers. It should still call the CLI commands documented here, but it owns the reasoning workflow and packet shape.
• lore-brief (lore:brief): use for daily/rolling continuity briefs, proposal-only synthesis, learned signals, memory-card candidates, and contradiction candidates. It must not create jobs, issues, wiki pages, skills, tasks, memory cards, code, or automations unless the user explicitly asks for that follow-up.
• lore-handoff (lore:handoff): use for compact continuation packets for the next agent, with verified/open/stale/risky sections, artifacts, proposals, memory-card candidates, contradiction candidates, and next actions.
• lore-dev-verification (lore:dev-verification): use when changing this repository, workflow skills, CLI/MCP retrieval, packaging, adapters, store compatibility, privacy, or destructive-memory behavior.

If a workflow skill reports that Lore is stale or unsynced, come back to this skill for the mechanical recovery command (lore sync, lore index, hook setup, or adapter work). If lore status --json shows schemaVersion greater than supportedSchemaVersion, read-only search may work but write recovery such as lore sync can be refused; update Lore before promising that sync will fix freshness.

The one rule that governs everything: drill down, never dump

A session can be thousands of messages — millions of tokens. Never pull a whole session into context. Every command here returns a bounded amount on purpose; your job is to narrow from a broad question to the few messages that actually answer it. The loop is always the same, and it is self-threading:

broad words  ──▶  lore search        ──▶  output hands you real ids
                                            (message <id>, session <id>)
                          │
        pick the id you want  ─────────────┘
                          ▼
   lore get <id> --full        (read that one message in full)
   lore context <id>           (see what surrounded it)
   lore session <id> --around <id>   (jump into that spot in the session)

You never invent an id. Every id you pass to get / context / session came verbatim out of a previous command's output. Searching is how you discover ids; the drill-down commands are how you spend them. If you catch yourself about to page through an entire session to "find" something, stop — that's what lore search --session <id> is for.

Lead with the narrowest query that could work. Broaden (drop filters, switch to --relevant) only if it comes back empty. Reading one message in full beats skimming fifty snippets.

Find — lore search

Keyword search (bm25) across every indexed session. This is the front door:

lore search "fts tokenizer"                  # keyword search, newest-ranked by relevance
lore search "fts tokenizer" --relevant       # blend keyword strength with recency
lore search "alamo" --json                   # { count, hits } — same envelope as MCP search_memory

Each hit leads with its message <id> and session <id> — those are the keys you drill with next. Snippets are elided; pull the full text with lore get <id> --full.

Filters (all optional, combine freely):

--relevant reranks by keyword strength × recency (1/(1+ageHours)), so a fresh memory beats an old one with the same words. It honors every filter above — including --session, which makes it a strict superset of the MCP find_relevant tool. Use it for "what did I do recently about X"; use plain search for "find the best keyword match, whenever it happened."

Read back — lore get, lore context

Once search hands you an id, read it:

lore get <message-id>                 # elided snippet (cheap)
lore get <message-id> --full          # the complete stored text
lore get <message-id> --json          # the message detail object (MCP get_message shape)

To see what was happening around a message — the question that prompted an answer, the reply that followed a decision:

lore context <message-id>                       # 5 before / 5 after, anchor flagged with >
lore context <message-id> --before 2 --after 8  # asymmetric window
lore context <message-id> --json                # { messages } (MCP get_context shape)

An unknown id returns { "error": "not_found", "message_id": "…" } and a non-zero exit — the same envelope MCP gives, so a script can branch on it.

Navigate — lore session, lore sessions, lore timeline

Skim or jump within one session. lore session walks a session's folded timeline (primary + subagent messages interleaved in real time) a bounded page at a time. It never dumps — a cursor carries you forward:

lore session <session-id>                          # first page (default 30, capped at 40)
lore session <session-id> --limit 40 --cursor <c>  # next page; <c> came from the prior page's output
lore session <session-id> --around <message-id>    # jump straight to a known message's neighborhood
lore session <session-id> --around <message-id> --before 10 --after 10

The human output trails each page with the exact --cursor command to continue. --json gives { messages, nextCursor } (paged) or { messages } (--around), matching MCP get_session. Prefer --around over page-walking: if search already gave you the id you care about, jump to it instead of reading from the top.

List recent conversations to find a session id when you don't have one:

lore sessions                                  # rollups, newest first
lore sessions --project <path> --source codex --limit 20
lore sessions --json                           # { count, sessions } (MCP list_sessions shape)

Each rollup shows message count, project, and time span. Copy a sessionId straight into lore search --session <id> or lore session <id>.

Chart activity over time — useful for picking a --since/--until window or spotting when a project was hot:

lore timeline                                  # message counts bucketed by day
lore timeline --bucket hour --project <path>   # finer grain, one project
lore timeline --json                           # { buckets } (MCP timeline shape)

Write memory in live — lore push

To add a session that has no on-disk transcript (a live process, a homegrown tool), pipe one normalized JSON batch to lore push:

echo "$BATCH_JSON" | lore push
cat batch.json | lore push

The batch is { sourceFile, messages, toolCalls } (the exact shape of the MCP push tool). It's validated whole at the boundary and written idempotently — re-pushing the same batch never duplicates. On success it prints the write result ({ sourceFileId, sessionId, messages, toolCalls }) and exits 0. On a malformed batch it prints { "error": "invalid_batch", "detail": "…" } and exits non-zero, so a caller can tell a rejected write from an accepted one. push is data only — it never receives or runs code.

For the record shapes, see src/core/records.ts; for when to push vs. write an adapter, see `references/setup/index.md` (PUSH vs PULL decision).

Remove memory — lore forget, lore exclude

These are destructive, irreversible commands. They follow a mandatory two-step confirmation: run the bare command to see the exact scope and counts; run again with --confirm to execute. The store is never touched on a bare run.

# Forget one session (point-in-time deletion -- future sessions still indexed)
lore forget --session <session-id>              # preview: shows message + tool-call counts
lore forget --session <session-id> --confirm    # execute: delete rows + write tombstone

# Find the session id first if you don't have it:
lore sessions --project "$PWD" --source claude-code --limit 10

# Forget all sessions for a project (future sessions still indexed)
lore forget --project /path/to/project          # preview
lore forget --project /path/to/project --confirm

# Exclude a project (delete existing + bar all future captures until removed)
lore exclude --project /path/to/project         # preview
lore exclude --project /path/to/project --confirm

# Manage standing exclusions
lore exclude --list                             # list all standing exclusions
lore exclude --remove /path/to/project          # lift an exclusion (does NOT restore deleted data)

Verb semantics:

• forget = point-in-time delete. What exists now is gone; future sessions for

the same project still get indexed.

• exclude = standing rule. Deletes existing data AND bars all future captures

from the project until --remove is called. Use when you never want a project in lore.

How tombstones protect you: After a forget or exclude --confirm, a tombstone is written to the store. Re-indexing or lore push with the same session or project id is silently dropped -- the data cannot come back through the normal ingestion path.

A mistyped id/path is safe. The preview for a non-existent session or project shows zero counts and exits 0. Nothing is deleted.

Human-in-the-loop rule (enforced for agents)

An agent using lore forget or lore exclude MUST follow this sequence and MUST NOT supply --confirm on its own initiative:

>

1. Run the bare command (no --confirm) and surface the exact scope and counts to the human. 2. Wait for explicit human approval of the displayed scope. 3. Only after the human confirms: re-run the same command with --confirm.

>

This is not optional ceremony. forget/exclude are irreversible. Showing the blast radius before acting is the safety contract. An agent that decides the counts look fine and adds --confirm on its own has violated this rule.

These commands are intentionally CLI-only and are NOT exposed as MCP tools. A compromised MCP client must never be able to wipe memory.

Index & teach new formats — see references/setup/index.md

This part of the skill is about using what's indexed. Getting transcripts into lore — lore setup (auto-detect Claude Code / Codex), lore index <dir>, lore sync codex (incremental active Codex catch-up), lore sample <dir>, lore hook (compaction capture for harnesses that emit a transcript path), and writing a reviewed code adapter for a brand-new harness (the checkAdapterConformance round-trip) — is the deterministic onboarding flow in `references/setup/index.md`. Read that reference when the answer is "the thing I want isn't indexed yet"; don't re-derive the procedure here.

Quick reference only:

lore setup            # detect known harnesses here, index, self-verify, print MCP registration
lore index <dir> [--source <name>] [--subagents] [--no-redact]
lore sync codex       # incremental active Codex catch-up for cron/launchd/manual use
lore sample <dir>     # summarize an unknown transcript dir's on-disk format

Understand the store (so you can extend it)

One SQLite file, queryable read-only while another process writes (WAL). Three ids, deliberately distinct:

• `source_file_id` — one physical transcript file (the ingest/watermark unit).
• `session_id` — one logical conversation, shared by a primary file and its

subagent files. This is what folds primary + subagent messages into a single timeline.

• `message_id` — a synthetic hash of (source_file_id + uuid + seq); uuid

alone collides across and within files, so it is never the key.

Tables: source_files, sessions (rollups: project, branch, first/last timestamp, message_count), messages (role, timestamp, project, branch, model, agent, skill, text, text_truncated), tool_calls (input/result/is_error, linked to their message_id), and a messages_fts FTS5 external-content table kept in sync by triggers. The tokenizer keeps _-. as token chars, so identifiers and paths (getUserById, foo.bar.ts, trust-metadata) are searchable as whole tokens — search for the symbol, not a fragment. Full schema: src/core/store/schema.ts.

LORE_DB overrides the store path (default ~/.lore/lore.db) — handy for pointing at a fixture store in tests.

CLI ⇄ MCP parity (why the server is unnecessary)

Every --json output is byte-shaped like the matching MCP tool's response, so "the CLI does what the server does" is verifiable, not asserted:

Worked recipes

"What did we decide about the FTS tokenizer?"

lore search "fts tokenizer" --relevant          # → hit with message <id>, session <id>
lore get <message-id> --full                    # read the decision in full
lore context <message-id> --before 3 --after 3  # see the reasoning around it

"Pull up my last session in this repo and find where we touched auth."

lore sessions --project "$PWD" --source claude-code --limit 5   # newest first → grab a session id
lore search "auth" --session <session-id>                       # only auth hits, inside it
lore session <session-id> --around <message-id>                 # jump to that spot, read forward

"When was this project active, and what happened in May?"

lore timeline --project "$PWD"                                  # spot the busy buckets
lore search "<topic>" --project "$PWD" --since 2026-05-01 --until 2026-06-01

Checklist

• [ ] lore help works (else see references/setup/index.md to install)
• [ ] Started from a search; used the ids it handed back (never invented one)
• [ ] Narrowed with filters / --around instead of paging a whole session
• [ ] Read the few right messages in full rather than skimming many snippets
• [ ] Did it all with no MCP server running