dsh-evolve

dsh-evolve

Self-evolving memory and skill lifecycle for DeepSeek Harness.

Your agent forgets everything between sessions. This plugin gives it durable memory, turns repeated procedures into reusable skills, and — crucially — keeps that knowledge from growing into a noise pile. Real evolution is mutation *plus* selection *plus* pruning; most memory plugins only do the first.

The plugin ships blank. It has no preloaded opinions about you or your work: only mechanisms and rules. Everything it learns is local to your install and never leaves it.

Requirements

Degradation is graceful by design: if SQLite/FTS5 is unavailable the plugin falls back to pure bigram recall, and any optional dependency that's missing disables only its own feature. It never blocks the harness from booting.

Install

Straight from this repository — no npm package needed:

dsh plugin --profile web add github:chenzheshushi-commits/dsh-evolve

Pin a specific release instead of tracking main:

dsh plugin --profile web add "https://github.com/chenzheshushi-commits/dsh-evolve/releases/download/v0.5.1/dsh-evolve-0.5.1.tgz"

Then restart the harness — tools are discovered at startup, not hot-reloaded.

Or clone for development:

git clone https://github.com/chenzheshushi-commits/dsh-evolve.git
cd dsh-evolve
pnpm install
pnpm run build      # builds the web-settings client bundle
pnpm run test       # smoke + registration probe + web-route e2e

What it does

Cross-session memory

Structured records (fact / preference / decision / lesson / todo / note) with scope (user = everywhere, project = here) and importance 1–3. Storage is JSON as the source of truth plus a Markdown mirror you can read and hand-edit.

Recall is zero-token and deterministic: bigram-Jaccard similarity fused with SQLite FTS5 BM25 through Reciprocal Rank Fusion. No embedding API, no per-turn model call. CJK text is tokenized correctly (searching 苹果 does not match 水果).

Relevant memories inject automatically each step based on the current message, and durable user preferences/facts inject as an always-on snapshot at the start of every turn.

Tiered approval, not "confirm everything"

Model-written memories pass through a deterministic gate that decides auto-confirm vs. hold for review, judged only on properties a model cannot flatter:

• reversibility (importance level)
• conflict with something you already confirmed
• overlap with existing memory
• whether the write traces back to something you actually said

Obvious, reversible, user-anchored writes land automatically. Risky or uncertain ones queue for review. The gate deliberately ignores the model-supplied kind field — letting a self-reported label decide its own exemption would be no gate at all. Auto-confirmed entries stay visible and revocable, and one config flag returns you to review-everything behavior.

Reinforcement: what you repeat gets stronger

Re-observing the same understanding doesn't duplicate it — it reinforces it. The observation count rises, importance climbs at a configurable threshold, and the better-quality phrasing is kept rather than blindly overwritten. Confidence is surfaced (low / medium / high) so the agent can weight established knowledge over one-off remarks.

Skills that improve instead of accumulating

High-value lessons sharing a tag crystallize into a SKILL.md. New evidence refines the existing skill in place — versioned, with your hand edits preserved — instead of spawning a near-duplicate.

Curation runs a real lifecycle: active → stale (idle N days) → archived. Archiving moves a skill out of the active catalog and is fully reversible. Nothing is ever deleted. A backup is taken before every mutating operation, so a bad refine or a hasty archive can be rolled back.

Anti-bloat convergence

The half most memory systems skip.

Skills: detects near-duplicate skills by content similarity and flags refinement-bloated files. Merging generates an umbrella skill and archives the originals (reversible). Folding compacts stacked refinement sections back into clean prose. Candidates that were never actually loaded rank first — duplicated *and* unused is the strongest case for merging.

Memory: a hard character budget that never silently drops anything (over-budget returns trim candidates for you to decide on), a gate against reworded near-duplicates and thin low-signal writes, and promotion of well-reinforced project memories to global scope.

Detection is always on and costs zero tokens. Every mutating action is opt-in.

Background review

At the end of a turn (throttled), an isolated LLM pass replays that turn's conversation snapshot and asks what's worth remembering. Suggestions route through the same approval gate — the reviewer proposes, it never writes directly.

It runs as a standalone call, so your main conversation and prompt cache are never touched, and because it's a plain text completion with no tools attached it is structurally incapable of side effects. It can be pointed at a different (cheaper or stronger) model than your main one.

Weak models degrade safely: a malformed review is skipped, so the worst outcome is "nothing learned this turn" — never "something wrong learned."

Knows you, and shapes tools to you

Confirmed user-scope preferences and facts accumulate into an auto-grown profile you can inspect, ordered by how consistently you've shown each one.

Skills can also carry a user-style overlay: a small instruction layer applied when the skill is used, derived from your profile. The underlying SKILL.md is never rewritten, so the overlay is fully reversible — clear it and the skill is vanilla again.

Maintenance sweep

A single tool aggregates every read-only check — archivable skills, merge candidates, bloated files, memory budget, promotion candidates, and whether enough outcome data has accumulated to be worth scoring — into one report. Safe to run on a schedule from an external cron; the plugin never installs an internal timer.

Tools

Memory: memory_remember memory_recall memory_index memory_confirm memory_confirm_batch memory_auto_review memory_profile memory_budget memory_promote memory_forget

Skills: crystallize_skill refine_skill skill_curator archive_skill restore_skill skill_rollback converge_skill fold_skill skill_style

Ops: evolve_maintain memory_stats skill_stats

Configuration

Everything is configurable through the plugin's settings page (web profile) or your DSH config. Notable switches:

The LLM is only ever used for optional auxiliary passes — skill refinement, background review, and skill merging. All of them are single-shot, skippable, and fall back to deterministic behavior on failure. Nothing runs in your main loop.

Design rules

• Never break the harness. Every failure path degrades quietly; the plugin cannot prevent a boot.
• Never delete user assets. Archive, back up, roll back — but never destroy.
• No internal timers. In-session work hangs off events; offline work is an external cron calling a tool.
• Ship blank. No preloaded personal data. What it learns stays on your machine and is never packaged.
• Mechanisms over model smarts. Safety comes from deterministic rules, so swapping models changes quality, never safety.

What's new in v0.5.1

Fix: the always-on preference snapshot now reaches every new conversation.

The Tier 1 always-on snapshot — the durable user preferences/facts block injected at the start of each turn — was deduplicated with a process-global last-key. Because durable preferences rarely change, the snapshot text stayed identical, so after the *first* conversation injected it, every later conversation's first turn was silently suppressed and never received the block at all. The per-step relevant-recall injector had the same class of cross-session leak on its repeat-suppressor.

• Dedupe is now per-session, keyed by the session via a WeakMap. Each new conversation gets the always-on block on turn 1; within a single conversation an unchanged snapshot is still skipped (the prompt-cache protection that dedupe was meant to provide is preserved). The WeakMap is reclaimed with the session — no manual cleanup, no leak.
• Regression test drives the real apply(ctx) with two independent sessions and asserts both the cross-session injection and the within-session suppression. Retrieval baseline unchanged (5/5 recall, MRR 1.0, R6 drift 0).

No config or API changes; no migration.

What's new in v0.5.0

Autonomy becomes a user-chosen dial, and Chinese retrieval is fixed at the root.

Earlier versions hard-wired how much the memory could decide on its own. v0.5.0 makes that a product setting, on both the ingest and the disposal side — deliberately asymmetric, because an ingestion mistake is an *addition* (visible) while a disposal mistake is a *subtraction* (invisible).

Ingestion autonomy — approvalMode (three tiers)

• manual — every model write waits for your confirm. balanced (default) — reversible writes that are anchored to a literal user utterance *or* near-duplicate of a confirmed memory auto-confirm; everything else is pending. autonomous — any reversible, non-conflicting write auto-confirms.
• autonomous still forces conflicts and high-importance (imp 3) memories to pending — the tier split sits *after* the conflict/importance scan, so it's a structural guarantee, not a fragile if.
• Bounded so it can't flood the store: at most reviewMaxAutoPerTurn auto-confirms per background-review turn (rest fall to pending), and a hard maxPendingQueue cap on the one region that can be losslessly refused. Confirmed memory is char-budget bound, pending is count-bound — neither pool grows without limit.
• Background review can no longer take the anchored auto-confirm shortcut on its own say-so (anchoredToUser is caller/store-derived, never model self-report).

Disposal autonomy — disposalMode (two tiers)

• manual (default) — nothing proposed automatically. suggest — when idle, recompute and surface low-value candidates for your review. Zero auto-deletion in any tier — heat stays a read-only ordering signal, physical deletion is never automatic; you still act on candidates through the two-stage prune panel.
• Candidate rule is objective and non-heat: never injected *and* never recalled (both channels zero) + past an explicit cool-off (disposalMinIdleDays), excluding pinned / protected-kind / pending / recent. Skills never enter any automatic tier (fold/archive stay manual). The tidy tier (bounded auto soft-delete) is deferred to v0.6.x alongside tombstone GC.

Retrieval (Chinese recall fixed)

• R1/R2 — tokenizer bug fixed. A greedy {2,} regex used to swallow an entire Chinese query into one token, so any multi-word paraphrase scored zero. Now runs match fully or fall back to down-weighted 2-gram fragments (stopword-filtered, capped), with a query-length-adaptive threshold. Recall up, precision held (adversarial false-match set stays at zero).
• R3 — tags fold into the FTS index, bridging part of the synonym gap at zero new dependency. R5 — retrieval degradation is now visible (fused vs bigram-only vs fts-degraded) instead of silently dropping quality. R6 — extended CJK ranges (Ext-A / Compatibility), verified to cause zero drift in the adjudicator's similarity thresholds on the real store (reproducible via pnpm run test:baseline).

Observability / audit

• Background review runs land in the JSONL audit. Pending records carry the source-context snippet they were drawn from. The prune preview is tabular.

All new config is conservative by default (balanced / manual) — existing behavior is unchanged until you opt in via the two new blocks on the settings page.

What's new in v0.4.2

The missing half of "self-evolving": the human-facing pruning page.

v0.4.0/v0.4.1 gave you the evolution loop (tiered approval, reinforcement, anti-bloat convergence, background review). v0.4.2 closes the loop on the *human* side — there was previously no UI to act on prune candidates, only back-end tools. You can now prune from the settings page:

• Soft-delete (reversible). Forgotten memories get a forgottenAt tombstone and disappear from recall / injection / crystallization, but stay in the store until you restore them. The MEMORY.md mirror gets a separate "## Forgotten (recoverable)" section so they never silently mix with active memories.
• pinned — three-tier protection. Pin a memory and it is locked from every code path: never enters prune candidates, never overwritten by near-duplicate reinforcement, never deleted without an explicit confirm=true. The protection lives in the data layer (one of the two places every delete goes through), so it holds regardless of whether the delete came from the panel, a tool call, or a future code path.
• Protected-kind review area. preference and decision memories are *not* direct-deleteable — the panel shows them in a read-only "Protected records (special review needed)" section rather than giving a button that does nothing.
• Heat is a read-only ordering signal. Each memory gets a power-law coldness score H = 1 / (1 + λ·Δt)^α. Time basis is accessedAt || createdAt — never updatedAt (merging / refining bumps updatedAt but that is not "access"; treating such a bump as decay would silently demote actively-used memory). Heat only orders prune candidates; it never archives anything automatically.
• Two-stage panel: preview → execute. Stage 1 (POST /prune/preview) builds an in-memory plan and returns a planDigest. Stage 2 (POST /prune/execute) consumes it. The plan registry uses atomic claim (synchronous consumed-flag flip *before* the applyPlan await) so double-click / retry / resend cannot re-execute — without it, skill-converge would create duplicate umbrella skills under load.
• Per-target ETag staleness check. Each target carries the etag it had at preview time. If something else mutates it before execute, that target is skipped (not-found / stale) with a reason; the rest of the plan still applies. No whole-plan failure.
• JSONL audit, fail-open + amortized ring-trim. Every run is appended to .evolve-audit.jsonl (500-row cap). The audit write is *fail-open* — a disk error warns, never blocks the prune.

A2 layout in the settings page: approval queue (existing) at top, then the new controlled-prune block (candidates + preview/execute + protected area + forgotten list), then overview below. Pinned rows render their checkbox disabled.

Excluded by design: local vector models, semantic search, knowledge graphs (too heavy for an optimization, not a rewrite). All four pure-logic mechanisms adopted — heat, JSONL audit, two-stage preview→execute with registry, idle refresh — were chosen because they add zero new dependencies and respect the "detect automatically, dispose explicitly" principle. The community is chenzheshushi-commits/dsh-evolve on GitHub; issue reports welcome.

License

MIT