Chronos MCP

The memory server that forgets responsibly — and can tell you what it used to believe.

Every output in this demo is real engine output — it's generated by scripts/make_demo_gif.py running the actual server code against a throwaway database.

Claude forgets everything when you start a new chat. Most memory servers fix that by hoarding: everything saved forever, every recall dumped into context, no notion of whether a memory is still true. Chronos takes the opposite stance — memory that behaves like memory:

• Time-travel: query_at reconstructs what Claude knew at any past timestamp, including content that has since been edited. "What did we know about the auth system on March 1st?" is a one-tool answer.

• Forgetting curves: every memory carries a confidence score and an FSRS retention model (the algorithm behind Anki). Unreviewed memories decay; confirmed ones stabilize. consolidate_memories runs a four-phase "dream" pass that merges duplicates, decays stale beliefs, and prunes what the system has lost faith in.

• Honest deletion: forget hides (auditable, reversible), purge destroys (content removed from every table, content-free audit stub left behind).

Everything is one local SQLite file. No accounts, no API keys, no cloud, no embedding-model downloads. Search is SQLite's built-in FTS5/BM25 with porter stemming, kept in sync with content by database triggers — the index can't drift, even on a crash.

Quick Start

1. Install

git clone https://github.com/Kodaxadev/ChronosMCP
cd ChronosMCP
pip install -e .

2. Add to Claude Desktop

Open claude_desktop_config.json and add:

{
  "mcpServers": {
    "chronos": {
      "command": "python",
      "args": ["/absolute/path/to/ChronosMCP/chronos_mcp.py"],
      "env": {
        "CHRONOS_DB_PATH": "/absolute/path/to/chronos.db"
      }
    }
  }
}

Claude Code (CLI):

claude mcp add chronos -- python /absolute/path/to/ChronosMCP/chronos_mcp.py

Restart Claude Desktop and you're done.

Related MCP server: Fuzzy Memory MCP Server

What You Can Ask Claude

Remember that we always squash commits before merging to main.

What do you know about our Git workflow?

What did you know about the auth system as of March 1st?

How confident are you in that memory about the rate limit?

Run a consolidation pass and show me what's gone stale.

Permanently delete the memory with my old API key in it.

Tools

Memory

Cognition

Graph (experimental, off by default)

An event-sourced project graph (tasks, dependencies, causal analysis over outcomes) ships behind CHRONOS_ENABLE_GRAPH=1. It is quarantined rather than deleted: the structural-similarity signal proved weak in practice (see docs/KNOWN_LIMITATIONS.md), and the core memory server is better without the surface area. Enable it if dependency-ordered task suggestions (suggest_next_tasks) are worth it to you.

Configuration

Optional: semantic search

By default search is lexical (BM25 + porter stemming) — zero extra dependencies, instant startup. If you want "car" to find a memory about an automobile:

pip install "chronosmcp[semantic]"     # local ONNX model, no API keys, no torch

then set CHRONOS_SEMANTIC=1 in the server env. Retrieval becomes hybrid: BM25 candidates and embedding nearest-neighbors (a ~70MB local model, downloaded once) are merged with Reciprocal Rank Fusion, so exact-term matches and meaning matches both surface. Results gain a semantic_similarity field. First startup on an existing database embeds your whole corpus — minutes, once; everything after is write-through and milliseconds.

Security Model — read this once

Chronos is a local, single-user tool. Its threat model is honest rather than reassuring:

• Plaintext at rest by default; SQLCipher opt-in. Without a key, the database is an unencrypted SQLite file — don't store credentials, and use purge_memory (not forget) when something must actually disappear. With pip install "chronosmcp[encryption]" and CHRONOS_DB_KEY set, the entire file (content, history, FTS index, WAL) is SQLCipher-encrypted. The PBKDF2 cost is paid once at startup, not per operation. Migrate an existing database with python scripts/encrypt_db.py encrypt. Know the trade: the key lives in an environment variable (readable by local processes running as you), and losing it means losing the data — there is no recovery.

• Memories are data, not instructions. Recall results carry a source field (user, claude, web, document). Content that originated outside the conversation should never be followed as a directive — a poisoned memory is a persistent prompt injection, and provenance is your tripwire. The tool descriptions instruct Claude accordingly.

• Injection-safe queries. Free text never reaches FTS5 MATCH syntax or SQL — queries are reduced to quoted terms, and every statement is parameterized.

• No network by default. The base server makes zero outbound connections — the only imports are mcp, numpy, and the standard library. The one exception is opt-in: CHRONOS_SEMANTIC=1 downloads its embedding model from Hugging Face once, after which inference is fully local.

How It Works

remember writes one row. A SQLite trigger indexes it into an FTS5 table in the same transaction — there is no in-memory index, no startup rebuild, and no window where the database and the search index disagree. recall ranks with BM25 (porter-stemmed), then re-ranks by an FSRS boost: memories that are confident and recently confirmed surface first, while stale unverified ones sink.

update_memory snapshots the old content into a version table before overwriting. query_at replays that history: it selects the memories that existed at the requested instant, resolves each one's content as of that moment, and ranks the reconstructed snapshot with an ephemeral BM25 index — so past and present searches score the same way.

The cognitive layer is the part that makes Chronos different over months of use. Memories you confirm get more stable (slower decay); memories that sit unreviewed lose confidence on every consolidation pass; memories that are both untrusted and faded become prune candidates. The system's beliefs converge toward what you actually use and verify — and every confidence change is written to an audit table, so you can always ask why it believes what it believes.

Out of the box, search is lexical (BM25 + stemming) — a deliberate zero-dependency trade-off. The optional [semantic] extra upgrades recall to hybrid lexical+vector retrieval with a local model (see Configuration); remaining boundaries are documented in docs/KNOWN_LIMITATIONS.md.

Development

pip install -e ".[dev]"
pytest tests/ -v        # 48 tests
ruff check .

CI runs the suite on Linux and Windows across Python 3.10/3.12/3.14. Architecture notes live in docs/ARCHITECTURE.md; the v3.3 → v4.0 redesign rationale is in CHANGELOG.md. The README demo regenerates with python scripts/make_demo_gif.py (needs pip install pillow) so it can never drift from real behavior.

License

MIT