Local Memory

Part of the [StudioMeyer MCP Stack](https://studiomeyer.io) — Built in Mallorca 🌴 · ⭐ if you use it

local-memory-mcp

Persistent local memory for Claude, Cursor & Codex. 25 tools. Hybrid retrieval (BM25 + vector cosine, RRF). Bi-temporal asOf queries + fact supersession. LLM-free contradiction detection + reflection. Portable JSON export/import. Multilingual embeddings. No cloud. No API keys.

Your AI assistant forgets everything when you close the chat. This fixes that.

Learnings, decisions, people, projects — stored in a single SQLite file on your machine that never leaves your computer. Built-in Knowledge Graph, duplicate detection, FTS5 keyword search, and (new in v2) hybrid retrieval that fuses BM25 with on-device vector cosine via Reciprocal Rank Fusion. The embedding model is multilingual (DE / EN / ES / 100+ languages) and runs locally — no API keys, no cloud.

📄 Deep dive: WHITEPAPER.md — architecture, design principles, the local-first contract, and an honest comparison vs. other memory systems.

Not affiliated with [`danieleugenewilliams/local-memory-releases`](https://github.com/danieleugenewilliams/local-memory-releases) — that is a different "Local Memory" project with the same descriptive name. This package is published as @studiomeyer/local-memory-mcp — always use the scoped name to disambiguate.

A note from us

We have been building tools and systems for ourselves for the past two years. The fact that this repo is small and has few stars is not because it is new. It is because we only just decided to share what we have built. It is not a fresh experiment, it is a long story with a recent commit.

We love building things and sharing them. We do not love social media tactics, growth hacks, or chasing stars and followers. So this repo is small. The code is real, it gets used, issues get answered. Judge for yourself.

If it helps you, sharing, testing, and feedback help us. If it could be better, an issue is more useful. If you build something with it, tell us at hello@studiomeyer.io. That genuinely makes our day.

From a small studio in Palma de Mallorca.

Quick Start

Claude Code

claude mcp add memory -- npx -y @studiomeyer/local-memory-mcp

Claude Desktop

With Node.js 22 installed, add the following to claude_desktop_config.json. This uses the published npm package.

{

"mcpServers": { "memory": { "command": "npx", "args": ["-y", "@studiomeyer/local-memory-mcp"] } } }

Cursor / VS Code

Add to .cursor/mcp.json or .vscode/mcp.json:

{

"mcpServers": { "memory": { "command": "npx", "args": ["-y", "@studiomeyer/local-memory-mcp"] } } }

Codex

# ~/.codex/config.toml

[mcp_servers.memory] command = "npx" args = ["-y", "@studiomeyer/local-memory-mcp"]

Automatic session tracking

You can make session tracking fully automatic so you never have to think about it.

Claude Code (CLAUDE.md): Add this line to your project's CLAUDE.md:

Always call memory_session_start at the beginning of each conversation and memory_session_end when done.

Claude Code (Hook): For a system-wide setup, add a SessionStart hook in ~/.claude/settings.json:

{

"hooks": { "SessionStart": [{ "hooks": [{ "type": "command", "command": "echo '{\"hookSpecificOutput\":{\"hookEventName\":\"SessionStart\",\"additionalContext\":\"Call memory_session_start now.\"}}'", "timeout": 5 }] }] } }

Both approaches make Claude call memory_session_start automatically. The CLAUDE.md way is simpler, the hook way works across all projects.

What it does

When you start a conversation, the server loads context from your last sessions so the AI knows what you were working on.

During the conversation, the AI stores patterns, insights, and mistakes via memory_learn. It records facts about people, projects, and tools via memory_entity_observe — building a knowledge graph over time. Every stored row is also embedded into a local 384-dim vector via the multilingual-e5-small model.

When you search, the unified memory_search runs hybrid retrieval: FTS5 with BM25 ranking is fused with vector cosine via Reciprocal Rank Fusion (RRF, k=60). That bridges vocabulary mismatches ("send" finds "publish"), works across DE / EN / ES / 100+ languages, and matches even when the query has no exact token overlap with the stored content. If the vector extension can't load on your machine, search transparently falls back to FTS5-only — nothing breaks, you just lose the semantic half.

The duplicate gatekeeper still prevents storing the same information twice.

Hybrid Search (v2.0.0+)

memory_search({ query: "...", mode: "hybrid" })   // default

memory_search({ query: "...", mode: "fts" }) // keyword only memory_search({ query: "...", mode: "vector" }) // cosine only

Architecture

• search_fts (FTS5, BM25) — keyword recall, the v1 path.
• embeddings (sqlite-vec vec0 virtual table, float[384]) — vector recall.
• Reciprocal Rank Fusion (k=60) combines the two when mode: "hybrid".
• Embeddings come from Xenova/multilingual-e5-small (Apache-2.0) via Transformers.js, q8-quantized (~30 MB cache). Model loads lazily on the first embed call; runs entirely on CPU.
• Auto-embed-on-insert covers learnings, decisions, and entity observations. Entities themselves are not embedded — their attached observations carry the semantic surface.

Multilingual. The default model is trained on 100+ languages with strong DE/EN/ES retrieval. Mixing languages in your stored data is fine — query in one language and the cosine half still surfaces relevant results in another.

Environment overrides

• MEMORY_EMBED_DISABLED=1 — force FTS5-only (e.g. air-gapped or corporate-proxy network).
• MEMORY_EMBED_MODEL=... — swap in a different Transformers.js feature-extraction model.
• MEMORY_EMBED_CACHE_DIR=... — override the Transformers.js cache location.
• MEMORY_EMBED_DTYPE=fp32|fp16|q8|q4 — model quantization (default q8).

Lifecycle + Reflection (v2.1.0+)

v2.1 closes the gap between "store a fact" and "manage a memory over time". The schema has carried archived, lifecycle_state, valid_from, and valid_to since v1, but no tool exposed them. Now four tools do.

Bi-temporal asOf — "what did I know on date X?"

memory_entity_open({ id: "...", asOf: "2026-04-15" })

Returns the entity plus the observation set whose validity window contained 2026-04-15. The filter is valid_from <= asOf AND (valid_to IS NULL OR valid_to > asOf). Accepts any format SQLite's datetime() recognizes: ISO 8601 (2026-04-15T00:00:00Z), SQLite-style (2026-04-15 00:00:00), or date-only (2026-04-15). Without asOf you get the legacy live-view (every observation with valid_to IS NULL).

Design choice — valid-time only, not full bi-temporal. SQL:2011, XTDB, Datomic offer two-axis bi-temporal (valid-time × transaction-time). We do valid-time only; transaction-time lives passively in created_at but isn't queryable as a separate axis. For a local AI-memory product the question is "what did the AI *know* about X on date Y" — that's valid-time. Full bi-temporal matters for regulated audit trails (insurance, banking) — if you need it, reach for XTDB.

Scale note. The asOf predicate wraps valid_from in SQLite's datetime() for format-robust comparisons, which means the planner can't use an index on the column directly. For corpora of <1000 observations per entity (typical) the scan is sub-millisecond. If you have an entity with 10k+ observations, add an expression index — CREATE INDEX idx_obs_valid_from_dt ON entity_observations(datetime(valid_from)) — and the predicate becomes sargable again.

Contradiction scanner — LLM-free

memory_contradictions({ minCosine: 0.75, minConfidenceDrift: 0.2 })

memory_contradictions({ entityId: "...", limit: 20 })

Surfaces observation pairs that are semantically very close (cosine similarity above minCosine) but disagree on either:

• negation marker XOR — one side asserts, the other negates (regex covers EN / DE / ES / Catalan).
• confidence drift — same surface claim, very different confidence values.

LLM-free on purpose — the no-API-key promise holds. The heuristic is conservative; the AI client judges. Pure duplicates (no negation, no confidence drift) are not flagged. The cosine math runs in SQL via vec_distance_cosine, so the extension must be loaded; on platforms where it isn't, the tool returns VECTOR_DISABLED with a clear message instead of degrading silently.

Calibration. Default minCosine = 0.75 follows 2026 retriever-tuning literature (SparseCL on Arguana; Milvus threshold-tuning guidance) which finds a sharp false-positive rise below 0.7. Lower to 0.6 for recall-heavy use, raise to 0.85 for precision-heavy. The negation regex covers EN / DE / ES / Catalan / Portuguese / Italian / French — the seven languages multilingual-e5-small handles strongest.

Supersede — retire a stale fact (v2.2)

memory_observation_supersede({ observationId: "..." })

memory_observation_supersede({ observationId: "...", supersededById: "..." })

The scanner *finds* a stale fact; memory_observation_supersede *acts* on it. It sets the older observation's valid_to (a tombstone) so the fact stops surfacing in live memory_search and memory_entity_open — but the row stays in the database, so an asOf query still returns it as the belief that was current before the cutoff. This is the Zep fact-supersession pattern: invalidate, never delete. Pass supersededById and the cutoff becomes the newer fact's valid_from (the old fact was true right up until the new one was recorded); pass validTo for an explicit instant, or neither for now. Same-entity and self-supersede guards keep the operation sane, and a second call is idempotent (already_superseded).

Archive + update — lifecycle for learnings

memory_learn_archive({ learningId: "...", reason: "wrong" })

memory_learn_update({ learningId: "...", content: "…", confidence: 0.9 })

archive is a soft delete: the row stays in learnings (with archived = 1, archived_at, and lifecycle_state = 'archived' | 'archived: '), the embedding stays in vec0 (so asOf-style cross-references can still resolve), but recall / search / the duplicate check all filter it out. Idempotent.

update edits a live (non-archived) learning. If content changes we re-embed in the F4 atomic pattern (compute outside the transaction, write i