Your AI agent forgets everything and it gets more expensive every week. Fix both.

Measured, not promised

The benchmark ships in the repo. Identical task sets, two strategies: naive transcript replay versus addressed recall over typed, content-addressed memory shards — measured with real API token counts on a pinned model (claude-haiku-4-5). Reproduce it: python -m bench.run --real

At Haiku pricing ($1/MTok in), a task at 1,000-decision history costs ~$0.140 under replay vs ~$0.003 with addressed recall — 47× cheaper, with 3.6× lower latency (4.5s → 1.2s). On Sonnet/Opus pricing, multiply the dollar gap by 3–5×. Model version is recorded with every figure; a zero-cost mock mode reproduces the scaling shape offline.

How it works

Memory is a Merkle DAG of four typed shards — facts, decisions, gotchas, artifacts — indexed by a constant-size manifest. An agent loads the manifest and resolves only the shards a task needs. History is append-only and hash-chained: every commit produces a root, every shard has an inclusion proof, and one flipped bit anywhere is caught on the next read. Works with any agent via MCP, HTTP sidecar, CLI, or the Python SDK — all four produce byte-identical state, verified in the test suite.

$ python -m cli --dir ./mymem recall --query '{"type":"gotchas","contains":"db"}'
{ "shards": [ { "key": "gotcha.sqlite-wal",
                "body": { "failure": "WAL mode dropped under network mount" } } ] }

$ python -m cli --dir ./mymem prove --claim '{"type":"gotchas","key":"gotcha.sqlite-wal"}'
{ "verified": true, "root_cid": "bafkreidd…" }

Free forever. Hosted when you want it.