Roadmap

What is done, and what is next. The measurements behind the benchmark entries are in the repository’s benchmarks/.

Done

A graph memory core as MCP tools plus skills, no model in the library.
Proactive hints, session-start context and background capture on hosts with hooks.
A CLI for every memory operation, alongside the MCP tools; background capture writes through it.
The periodic pass, memoose maintain, offered daily on hosts with hooks.
Procedural memory after the Procedural Graphs paper: transitions with condition, advice and pitfall, guidance keyed on a declared position, outcomes counted on every transition a session took.
memoose install puts the skills, hooks and the memory-keeper agent into a host; the MCP server is opt-in.

A live test of the procedural loop. The unit tests cover guidance, positions and outcomes; nothing yet shows an agent declaring its position often enough for the trace to be useful, or following the guidance. The plan is a handful of real tasks on this repository run with and without declared positions, outcomes recorded, counters read back.
Propose an eval for maintained memory. No published benchmark asks whether a body of facts stays trustworthy as it changes: a fact was revised, two sources disagree, where did this come from. An eval only its author runs is not evidence, so this needs a design other systems can run before any number from it is worth publishing.
A token-cost benchmark. Accuracy is reported everywhere and cost almost nowhere, though cost is what a memory system charges you per turn. No comparable methodology across systems exists yet.
First PyPI release: push the v0.4.0 tag; the publish workflow does the rest.
Exercise the delegated subagent path live on a host without hooks, and a live end-to-end run of maintain, the way capture was verified. Re-check capture quality too: the capture prompt changed shape when it moved to the CLI and has not been measured since.
An ANN index, if a project ever passes about 200k vectors. SQLite itself has room well past that.

A hosted service, a UI, multi-user access control, other storage backends, or winning LoCoMo. The vision page says why.