Agent Memory

Description

Agent-driven long-term memory for DeepSeek Harness: scoped memory (global + per-workspace), layered entries (fact/knowledge/episodic), **role dimension for multi-agent workbench** (role-stamped writes + role-policy retrieval admission), **value auditor** (read-only memory_audit + proposal log), **context-centroid injection** (auto-recall query = weighted recent-turns terms, not just the last message + hit-rate metrics), **compaction-pipeline evidence layer** (per-bucket verdicts `not-ended`/`no-sources`/`already-summarized`/`pending` + `<DSH_HOME>/memory-compress-trace.jsonl` + memory_health readout), time-bucket compaction (day→week→month→year), associative recall (related chains) + memory_relate navigation, auto recall injection (CJK bigram search, tail append) with **v0.10 quality gates** (session-level entry dedup + IDF term weighting + generic-word gate + min-top-score gate — 无相关记忆时静默不注入), **v0.11 evidence-access & storage-hygiene fixes** (recall/memory_browse render now surface entry ids — they were long present in the output schema but dropped at the render layer, leaving update/forget/memory_merge without an entry point; compaction checkpoints now key on a **day-granular title** so repeated compactions in one day merge into a single episodic entry instead of leaving parallel snapshots; memory_audit gained `summary_max_chars` so an over-budget summary is flagged even below the generic 12k threshold), agent-decided content.