mneme

The three flaws every AI assistant has

Mneme fixes all three at the architecture level — not the prompt level.

The killer feature — compaction resilience

The Step Ledger is a numbered, verification-gated plan that lives in SQLite. Every step records its acceptance check. When compaction wipes Claude's working memory, the next turn auto-injects a ~5K-token resumption bundle with the verbatim goal, completed steps, current position, and remaining work. No other MCP does this.

Benchmarks vs code-review-graph

Measured against the current state-of-the-art code-graph MCP.

With tree-sitter alone, vs. with mneme

Tree-sitter is a parser library, not a code-graph or MCP server. This table is what you get out of the box with raw tree-sitter against what mneme gives you by wiring tree-sitter into a persistent SQLite graph with 47 MCP tools.

How to reproduce

Every mneme number above is from benchmarks/ in this repo or directly measurable in source. To get your own numbers:

## 1. Clone and build
git clone https://github.com/omanishay-cyber/mneme && cd mneme
cargo build --release -p benchmarks --bin bench_retrieval

## 2. Index the repo you want to measure
cargo run --release -p cli --bin mneme -- build /path/to/repo

## 3. Run the full benchmark suite on that repo
./target/release/bench_retrieval.exe bench-all /path/to/repo \
    > benchmarks/results/$(date -I).csv \
    2> benchmarks/results/$(date -I).stderr

## 4. Individual benches (cold, warm, incremental, viz-scale, recall, compare)
./target/release/bench_retrieval.exe bench-first-build  /path/to/repo --format json
./target/release/bench_retrieval.exe bench-incremental  /path/to/repo --format json
./target/release/bench_retrieval.exe bench-recall       /path/to/repo benchmarks/fixtures/golden.json --format json
./target/release/bench_retrieval.exe bench-token-reduction /path/to/repo --format json

## 5. CRG head-to-head (requires Python 3.10+)
cd benchmarks/crg-compare
python -m venv .venv
.venv/Scripts/python.exe -m pip install code-review-graph
python run_crg_bench.py

Raw results from mneme's own run: benchmarks/results/2026-04-23.csv + benchmarks/results/crg-2026-04-23.json. Full methodology and caveats in BENCHMARKS.md.

Works with everything

One mneme install configures every AI tool it detects.

20 expert skills + 4 workflow codewords

Mneme ships a full skill arsenal. Claude auto-loads the right expert for the task. Four single-word verbs switch how Claude engages.

Codewords - one-word session control

Fireworks skills - auto-dispatched by keyword

Each skill is a full package: SKILL.md with triggers + protocol, plus a references/ folder of deep how-to docs. Skills sleep until a matching keyword appears in your message - a Rust task never fires the React skill.

Type firestart on your next complex task. Mneme loads the full arsenal, primes your code graph, builds a numbered ledger, and works through it with verification gates - one step at a time.

Architecture — multi-process, fault-tolerant, 100% local

Marketplace plugin (global / user / project scope)
└─ SUPERVISOR (Rust, Windows service / launchd / systemd)
   ├─ STORE           22-layer SQLite sharded per project + WAL
   ├─ MCP server      Bun TS, 47 tools, JSON-RPC stdio
   ├─ PARSERS         Tree-sitter, 27 languages, num_cpus×4
   ├─ SCANNERS        Theme / security / a11y / perf / drift
   ├─ MD-INGEST       Drinks every .md like CLAUDE.md
   ├─ BRAIN           Pure-Rust embeddings + Leiden clustering
   ├─ MULTIMODAL      Python sidecar — PDF / Whisper / OCR
   ├─ LIVE BUS        SSE/WebSocket push channel
   ├─ VISION          14-view WebGL desktop+web app
   └─ HEALTH          60 s self-test, SLA dashboard