Swapnanil Saha

An AST-aware codebase indexer with hybrid semantic + BM25 search, a symbol graph, and a persistent working-memory tier delivered through hooks — findings survive context compaction and session restarts instead of dying with the conversation. Any MCP-compatible AI code editor connects automatically. v1.1 added team mode — many agents, one shared memory; v1.3 added per-memory triggers, so a note resurfaces exactly when its path, symbol, or lifecycle condition matches; v1.5 added automatic episode capture, where the harness records what was tried and what failed, and distils the recurring arcs into notes; v1.8 added a machine-checked user-stated trust class, where a note quoting the user's own words verbatim proves it; v1.10 made expiry a state transition instead of a delete, so an aged-out note is marked expired and stays auditable rather than vanishing, added pinned notes that recall returns whatever the query says, and rendered working memory to a read-only MEMORY.md a person can actually open and read; v1.11 lets a note be retired after the fact without the "proven wrong" framing revocation carries, records whether the agent actually read the file a note is anchored to rather than assuming it, and opens working memory in your own editor while still writing every change back through the append-only lifecycle.

v1.11 separates being out of date from being wrong: a note can now be retired after the fact in favour of a newer one, without the "proven wrong" framing revocation carries, and reinstating one actually reverses it instead of leaving it half-retired. An anchor now records whether the agent ever actually read the file it points at, so a note anchored to a file nobody opened no longer looks identical to one that was genuinely bound. And working memory opens in your own editor, with every change written back through the append-only lifecycle rather than around it. v1.10 made memory expiry a state transition rather than a delete: an aged-out note is marked expired and keeps its place in the audit trail, directives never age out at any age, and a pinned note comes back from recall whatever the query says. It also rendered a workspace's notes to a read-only MEMORY.md, so what the agent remembers is finally something a person can open and read. v1.8 shipped a machine-checked user-stated provenance class (a note binding a verbatim excerpt of the user's words earns a stronger recall frame; the class is earned by a deterministic check, never claimed), after 1.6 and 1.7 made proactive delivery the default and made it worth reading. v1.5 shipped automatic episode capture with arc distillation, and a note lifecycle where a wrong note is revoked rather than deleted, so the mistake is not silently re-derived, on top of 1.3's per-memory trigger engine and 1.1's team mode. The research underneath it is now published: Delivery, Not Storage (arXiv:2607.20972), a controlled study of agent memory under repeated context compaction across 108 and 138 forced boundaries, graded token by token. A second report from the same programme is published beside it: Deterministic, Reproducible, and Wrong, on how our own deterministic scorer manufactured an agent-memory headline that the transcripts refuted, and the contradiction guard that would have caught it. Running alongside it: an upstream feature request to Claude Code (#78795) proposing triggered injection for auto-memory files, so the findings become platform behavior rather than a workaround.