A compiler for what your coding agent sees

What it does It’s a local-first Rust CLI that turns repository history — decisions, fixes, traces, tests, prior agent work — into typed context packs before inference. Durable repo knowledge goes into a byte-stable, blake3-hashed prefix; volatile task state goes into a fresh suffix; stale docs, contradictions, and secrets get excluded outright. Everything lands in a SQLite vault under .memory.cpp/, fully offline, with a ledger recording exactly what was sent and why.

The interesting bit The framing is the trick: this is a compiler, not a retrieval system. Facts pass through a MemoryIR with source anchors, trust, freshness, and sensitivity metadata, then optimizer passes score them by utility-per-token (volatility counts as cost). Negative memory — remembering what not to include — is a first-class feature, which is rare and genuinely sensible.

Key highlights

• Byte-stable cacheable prefix means the same repo state produces the same prefix across tasks, so provider prompt caching actually hits
• Freshness proofs tie memories to files, commits, or tests; unsupported memories are rejected before the prompt
• Rehydratable mem://file/...#L10-L50 handles ship compact facts and defer exact code until needed
• CAMC-Bench ablation: full pipeline sends 266 tokens vs 638 raw, blocks all stale/contradiction/secret leaks, passes the task oracle where raw, grep, BM25, and embeddings all fail
• Dogfooded on its own repo: 854 tokens vs 7,961 for pasting the docs (~89% smaller)

Caveats

• The benchmark corpus is synthetic — it measures the mechanism, not real-world win rates; the README says so plainly, to its credit
• It does not compress a provider’s KV cache; it reduces KV pressure by keeping tokens out of the prompt, and the README is explicit about this distinction
• The command surface is enormous (compile, prove, autopsy, fuzz-context, poison…) — expect a learning curve

Verdict Worth a look if you run long-horizon coding agents and are paying for repeated prefill or debugging stale-context patches. If you just want a memory store your agent can query, simpler tools exist — this one is for people who care about cache layout and provenance.