Status & Roadmap

August 20, 2026 · View on GitHub

中文 →

A reorganized development diary: what is implemented today, what is not, and where this can grow. Version 0.1.0, schema v5.

Iteration history (condensed)

The design was worked out over a series of numbered design notes (001–010 in the original memory-arch series) plus external architecture reviews:

  • 001–003 — design decisions, evidence/experiment, and the service roadmap that settled the two-memory model and the earned-trust lifecycle.
  • 004 — first implementation status: diary + store + service + the initial tool surface.
  • 005 — the consolidated final design around 生·用·修·记, which the implementation follows.
  • 006–010 — incremental rounds adding: experience lifecycle hardening, the ledger, the workbench UI, round-2 (fact dedup + corroboration, concern background, the thinking kind, and profile-snapshot injection), landing at schema v5.

✅ Implemented

Core store & lifecycle

  • Process-global SQLite store (memory.db) shared by all sessions; additive schema migrations (now v5).
  • Diary: append-only raw entries (memory_fact).
  • Extraction: memory_extract distills diary into facts and concerns.
  • Experience lifecycle: memory_ingest, memory_report, memory_revise, memory_refine, memory_verify, rollback; trust is earned by reported use, promoted/demoted by evidence.
  • Recall: memory_recall — keyword/situation relevance over the global library, with opt-in context narrowing.
  • Ledger: append-only audit trail (memory_ledger, memory_verify).
  • Consolidation: memory_consolidate for periodic housekeeping.

Semantic memory (facts & concerns)

  • Nine user-centric fact categories.
  • Fact dedup: same category+key+value merges into one entry with an incremented corroboration counter and merged source-diary ids.
  • Concern kinds: todo / thinking / idea / question / decision / commitment / other.
  • Concern background scene (schema v5) so the agent knows why a concern matters.
  • Concern status + tree structure; filtered listing and counts.

Integration & surface

  • Profile-snapshot injection: top facts + open concerns injected into the system prompt as a dedicated section.
  • Browser workbench: Fact Diary, Experiences, Ledger, and Human Ops pages with filters and paging.
  • Human Ops: manual promote / demote / pin / edit / delete, recorded in the ledger.
  • Remote API: a typert remote surface so the browser half talks to the host store.
  • memory-extraction skill: bundled guidance teaching the agent when/how to extract well.

Packaging (this repo)

  • Distributable package dsh-daoing-memory with a dsh.bundle.patch for automatic profile wiring.
  • Prebuilt host + browser artifacts shipped in lib/ (install performs no build).
  • One-command skill placement helper (scripts/install-skill.mjs).

🚧 Not implemented (yet)

  • Embedding-assisted recall. Recall is keyword/situation relevance only; there is no vector index.
  • Cross-agent / multi-tenant isolation. The store is one process-global library.
  • Automatic forgetting. Decay/forgetting is explicit (human or consolidation), not policy-driven.
  • Concern auto-resolution. Concerns don't yet close themselves when the underlying todo/question is settled.
  • Semantic dedup for experiences. Fact dedup exists; near-duplicate experiences are not merged automatically.
  • Fully decoupled standalone build. Artifacts are produced with the DSH build toolchain and shipped prebuilt; the repo does not yet rebuild from source on its own (see BUILDING).

🧭 Extension directions

Ordered roughly by leverage:

  1. Embedding-assisted recall — augment (not replace) keyword relevance with a vector index for fuzzy matching, keeping the trusted-library governance.
  2. Fully standalone build — vendor the build tooling so the package rebuilds from source without the DSH monorepo, enabling true independent iteration.
  3. Consolidation policies — decay, cold storage, and compaction heuristics for long-lived memory.
  4. Concern lifecycle automation — detect when a concern's loop is closed and propose resolution.
  5. Experience semantic dedup — merge near-duplicate experience families.
  6. Multi-tenant scoping — per-user or per-agent isolation on top of the global store.
  7. Marketplace maturity — publish to npm, list on the DSH plugin marketplace, CI build/release.
  8. i18n & more fact categories — broaden the user-centric taxonomy.

Versioning

  • 0.1.0 — first distributable release (this repo). Schema v5.

See CHANGELOG.md for release notes.