Память

Project memory for coding agents on DeepSeek Harness: similar past edits surface as precedents, out-of-scope edits are blocked, and cognition persists across sessions (47 upstream-verified fixes, gold-free 20-task run).

PLUR memory rendered into the system prompt on each assembly rather than exposed as a tool call, so the block is replaced instead of appended and context stays flat across a session. Fully local hybrid search (BM25 + BGE fused with RRF), plain YAML storage you can edit, per-workspace scoping, and a /plur-memory browser.

Hana-style assistant manager for DSH: visual preset management (cards, yuan personalities, avatar cropping) plus a full Hana memory stack (rolling summaries, layered compilation, SQLite fact store, deep memory).

Connect Trilium notes to the DSH agent via ETAPI with 20 agent tools: durable memory (remember/recall + auto-injected index), note management (CRUD/clone/attributes/revisions), full-text search, weekly-report workflow, attachments, calendar notes, backup/import, and a settings card.

LLM-driven remember/forget memory: after each turn a model decides what to keep (MemoryDirector), injects relevant memories before each step, dedups, and persists cross-session.

Compaction backend replacing LLM summarization with a deterministic semantic extractor (keeps code/paths/commands, drops chatter) plus 28.4x KV-compression accounting.

Organize folders and sessions into projects with shared PROJECT.md memory and AGENTS.md instructions injected into every session, plus a project tree sidebar.

Session context triage for long DeepSeek Harness sessions: deterministic screeners trim stale, duplicate, failed, and oversized content through the official ctx.compaction seam.

Cross-session event-cue memory for DeepSeek Harness: auto-extracts task/conversation cues and user language habits from every turn into a persistent global cue bank, then re-injects remembered context on topic-switch detection (keyword or vector matching).

Native per-workspace memory on the harness's own seams: facts and a bounded always-on profile on the storage-domain JSON unit, approval-gated writes with `(sessionId, seq)` citations, deterministic recall plus session-query FTS over past sessions — no external server, no custom SQLite.

Project-scoped persistent key-value memories with set/get/list/delete/search model tools and JSON-file storage.

Memory protocol enforcement plugin: bridges the opencode-memory MCP server, forces memory_weave before tool calls, auto-ingests turns, and injects memory context into agent steps.

PostgreSQL-backed durable memory for DSH agents: MCP memory tools plus a Web panel for reviewing memories and candidates.

Focus board for DeepSeek Harness agents: durable, model-maintained notes in the session workspace that pin the objective, constraints, and decisions across compaction and sessions, with automatic context injection, archive on clear, and an optional web panel.

Память между сессиями для модели: полнотекстовый поиск по всем прошлым сессиям (SQLite FTS5 через ctx.sessionQuery) и возврат наиболее релевантных фрагментов в текущий контекст.

Расширенные инструменты memoria: внедрение контекста сессии (профиль + извлечение), недавние решения, отчёт о состоянии, allow-лист пространств имён, граф связей памяти и поиск сущностей; дополнение к dsh-memoria.

Страж дрейфа предпосылок после сжатия: внедряет разовое уведомление, когда сводка сжатия теряет критический буквальный якорь.

Рабочая память на основе файла: дословно запоминает/извлекает ключевые предпосылки в файле заметок сессии, чтобы они переживали сжатие контекста без потерь.