GitHub - janbjorge/rekal: Long-term memory for LLMs. MCP server backed by hybrid search in a single SQLite file.

Long-term memory for LLMs. One SQLite file, no cloud, no API keys.

rekal is an MCP server that gives AI coding agents persistent memory across sessions. Memories are stored locally in SQLite and retrieved with hybrid search (BM25 keywords + vector semantics + recency decay). Nothing leaves your machine.

Works with any MCP-capable agent: Claude Code, Codex CLI, OpenCode.

Session 1:   "I prefer Ruff over Black"  → memory_store(...)
Session 47:  "Set up linting"            → memory_search("formatting preferences")
                                          ← "User prefers Ruff over Black" (0.92)
                                          Sets up Ruff without asking.

Install

pip install rekal
# or
uv tool install rekal

Requires Python 3.11+. On first run, rekal creates ~/.rekal/memory.db.

Setup — Claude Code

Three steps: add the MCP server, install the plugin, and disable built-in memory.

1. Add the MCP server — gives Claude Code the memory tools:

claude mcp add rekal rekal

2. Install the plugin — teaches Claude Code when to use those tools, and prevents conflicts with built-in memory:

claude plugin marketplace add janbjorge/rekal
claude plugin install rekal-skills@rekal

3. Disable built-in auto memory — add "autoMemoryEnabled": false to ~/.claude/settings.json:

{
  "autoMemoryEnabled": false
}

Why is this required? Claude Code's instruction priority is system prompt > CLAUDE.md > MCP server instructions. Built-in memory lives in the system prompt and always wins — without disabling it, the agent ignores rekal and writes to a flat file with no search, no deduplication, no ranking. The plugin's SessionStart hook replaces the context injection auto memory normally provides, so you don't lose anything.

What if I forget? The plugin's block-memory-writes hook will catch and block MEMORY.md writes as a safety net, but the agent wastes turns hitting the block. Disabling auto memory is cleaner.

Can the plugin do this automatically? No — Claude Code doesn't allow plugins to modify user settings. This manual step is the only way.

What the plugin provides

Hooks (automatic, no user action needed):

Skills (user-invocable):

Setup — Codex CLI

One step. rekal is a standard MCP stdio server — no plugin system, no competing memory to disable (Codex memories are off by default).

Add to ~/.codex/config.toml (Codex MCP docs):

[mcp_servers.rekal]
command = "rekal"

# optional: scope all memories to a project automatically
[mcp_servers.rekal.env]
REKAL_PROJECT = "my-project"

Instruct the agent to call memory_build_context at session start. Add to your project's AGENTS.md:

Call memory_build_context with your current task before exploring the codebase.

If you have enabled Codex memories (memories = true in ~/.codex/config.toml): disable them to avoid competing memory instructions.

[features]
memories = false

Setup — OpenCode

One step. OpenCode has no built-in memory system — rekal plugs in cleanly with no conflicts.

Add to opencode.jsonc in your project root (OpenCode MCP docs):

OpenCode does not auto-read AGENTS.md — you must list instruction files explicitly (OpenCode config docs). Add to your opencode.jsonc:

Tools

rekal exposes 21 MCP tools grouped into four categories.

Core — read and write memories:

Smart write — manage knowledge over time:

Introspection — explore what's stored:

Conversations — track session threads:

How it works

Storage

Everything lives in ~/.rekal/memory.db. Three subsystems share it:

• memories table — content, type, project, tags, timestamps, access counts, plus tier (durable or scratch) and optional expires_at
• FTS5 virtual table — full-text index over content+tags+project, auto-synced via triggers
• sqlite-vec virtual table — 384-dimensional vector index for semantic search

Memory links (supersedes, contradicts, related_to) are stored in a separate table. memory_supersede writes the new memory and creates a supersedes link in a single operation — old knowledge stays queryable with explicit lineage.

Tiers. Durable memories live forever; scratch memories carry an expires_at and are hard-deleted on server start once past their TTL. Search, timeline, and topics hide expired scratch entries automatically. Use scratch for in-flight hypotheses and working notes that should not pollute the durable store.

Embeddings

rekal uses fastembed with BAAI/bge-small-en-v1.5 (384 dimensions). Runs locally via ONNX — no API calls, no network. The model downloads once on first use (~50MB) and is cached.

Search

Every memory_search runs two parallel lookups, merges candidates, then scores:

score = w_fts × sigmoid(-BM25)                       ← keyword relevance    (default 0.4)
      + w_vec × (1 - cosine_distance)                 ← semantic similarity  (default 0.4)
      + w_recency × exp(-0.693 × days/half_life)      ← recency              (default 0.2, 30-day half-life)

Why three signals? Keywords miss synonyms ("deploy" vs "ship to prod"). Vectors miss exact identifiers. Recency alone buries important old knowledge. The blend covers all three failure modes.

Configurable weights. All weights and half-life are configurable at three levels:

Layers resolve per-key independently. A .rekal/config.yml setting w_fts and a DB override for half_life combine — each key uses its highest-priority source.

# .rekal/config.yml
scoring:
  w_fts: 0.6
  w_vec: 0.3
  w_recency: 0.1
  half_life: 14.0

Why SQLite?

• Single file — copy, back up, version-control, or delete to start fresh
• Zero config — no daemon, no port, no connection string
• FTS5 built-in — BM25 ranking without an external search engine
• sqlite-vec extension — vector search in the same process, no separate vector DB
• Sub-millisecond — local disk I/O, no network round-trips

Troubleshooting — Claude Code

Agent still writes to MEMORY.md

1. Check autoMemoryEnabled is false in ~/.claude/settings.json
2. Check the plugin is installed: claude plugin list should show rekal-skills

Agent doesn't call memory_build_context at session start

The SessionStart hook injects a reminder. If the agent ignores it, add to your project's CLAUDE.md:

Call memory_build_context before exploring the codebase.

Memories not being stored

Check the MCP server is running: claude mcp list should show rekal. If missing:

claude mcp add rekal rekal

Updating the plugin

Claude Code's plugin system may serve a stale cache after plugin install. If hooks or skills are missing after an update, clear the marketplace cache first:

rm -rf ~/.claude/plugins/marketplaces/rekal
claude plugin marketplace add janbjorge/rekal
claude plugin install rekal-skills@rekal

Architecture (for contributors)

Plugin (hooks + skills)
  │
  ├── hooks/
  │   ├── handlers/session-start.py       ← SessionStart: inject context reminder
  │   └── handlers/block-memory-writes.py ← PreToolUse: block MEMORY.md writes
  │
  └── skills/
      ├── rekal-init/    ← /rekal-init: bootstrap project knowledge
      ├── rekal-save/    ← /rekal-save: end-of-session capture
      ├── rekal-usage/   ← /rekal-usage: operational guide for tools
      └── rekal-hygiene/ ← /rekal-hygiene: maintenance

MCP Server (rekal)
  │ stdio (JSON-RPC)
  │
  mcp_adapter.py          ← FastMCP server, lifespan, instructions
  │
  ├── tools/core.py       ─┐
  ├── tools/introspection.py│─ thin @mcp.tool() wrappers
  ├── tools/smart_write.py  │
  └── tools/conversations.py┘
                            │
                    sqlite_adapter.py ← all SQL lives here
                            │
                            ├── SQLite (memories, conversations, tags, conflicts)
                            ├── FTS5 (full-text index)
                            └── sqlite-vec (vector index)

Instruction flow (single source per concern):

CLI

rekal serve    # Run as MCP server (default)
rekal health   # Database health report
rekal export   # Export all memories as JSON
rekal prune    # Bulk-delete memories by scope (dry-run unless --yes)

rekal prune requires at least one filter: --project NAME, --memory-type TYPE, --older-than-days N, or --before "YYYY-MM-DD HH:MM:SS". Without --yes it only reports the match count.

License

MIT