Long Term Memory

🌐 Language: English | 中文

Long-Term Memory

KinetAios auto-extracts "durable facts about the user" from each turn, stores them in SQLite, and injects them into the next turn's prompt. Cross-engine, cross-session.

Flow

turn N completes
  ↓
extractMemories(): calls LLM once, extracts durable facts from this turn
  ↓
inserts into the memories table (SQLite)
  ↓
turn N+1 starts
  ↓
memoryBlock(): recall-based injection (embedding cosine → FTS5 → recent-N)
  ↓
Direct: as the history[0] user message (_memory marker)
Claude Code: --append-system-prompt
Codex: prepended to the prompt

Runs in the background; does not block user input.

Extraction (extractMemories)

Around src/main/TaskManager.ts:282. Triggered on each done event:

• Take this turn's user prompt + assistant answer
• Call LLM with a small standalone prompt (does NOT enter directHistory)
• The prompt guides the model: only extract "durable facts about the user" — preferences, skills, long-term goals, fixed constraints. Not one-off task details ("write me a demo this time" doesn't count).
• Each returned fact is inserted separately into the memories table

Uses the current session's model (same engine, same model). GLM extracts via GLM, Claude via Claude.

Storage (SQLite)

src/main/store.ts:

CREATE TABLE memories (
  id TEXT PRIMARY KEY,
  content TEXT NOT NULL,
  conversation_id TEXT,           -- the session that extracted this; NULL = global
  created_at INTEGER NOT NULL
);

Each memory carries its source session (for "current channel" filtering).

Injection (memoryBlock)

src/main/TaskManager.ts:443:

Since v1.9.0, memoryBlock uses recall-based injection instead of full-set injection:

memoryBlock(conv)
  → build query from last 1-3 user messages (≤500 chars)
  → recallForInjection(query):
      1. embedding cosine (score > 0.25, top-15, ≥3 hits) → touchMemoryUsed
      2. FTS5 LIKE over memories table (≥2 hits)
      3. recent-N fallback (newest 15)
  → shellSafeMemory filter → inject as "## About the user" block

Why recall instead of full-set: with hundreds/thousands of memories, injecting all of them wastes tokens and dilutes relevance. The recall pipeline selects the 15 most semantically relevant memories for the current conversation — ~500 tokens instead of ~2000.

Direct engine special handling (v1.0 refactor)

memoryBlock is not concatenated into systemPrompt — it goes in as the history[0] user message (marked _memory: true). Reasoning:

• Anthropic's cache_control is set on the entire system
• When memory is in system, every memory change (a new fact extracted) busts the entire base+rules+context system cache
• Splitting it out keeps system stable across turns, so the cache hits
• Memory invalidation only affects that one small message

trim / compact always preserve _memory messages; on return, dropTransient filters them out and does not write back to directHistory (re-injected fresh next turn, prevents stale accumulation).

See Direct-Engine.

🧠 panel

Sidebar 🧠 button → long-term memory panel:

Action	Effect
Scope toggle: Current channel / All	Filter by source conversation_id
Inline edit	Edit content directly, save to DB
Delete	Remove a single entry, with confirmation

A main-window modal, not full-screen.

Import / export

⚙ → Long-term memory:

• Export JSON — writes to a user-chosen path. Structure: { version: 1, exportedAt: number, memories: Memory[] }
• Import JSON — accepts the above structure or a plain string[]. Dedupes by content, skips existing. Returns { imported: N, skipped: N }.

Good for: machine migration, backup, sharing memories across providers.

recall_memory tool vs long-term memory

Don't confuse them:

	recall_memory tool	Long-term memory
Source	history table (FTS5-indexed conversation text) + memories table (embedding cosine)	memories table (recall-based injection)
Trigger	Model invokes explicitly	Auto-injected every turn (top-15 by relevance)
Use case	"How did we solve X last time?"	"Who is this user, what do they like"
Volume	Every conversation	Only durable facts

recall_memory details: Tools-and-MCP.

Known limitations

• Extracts every turn: one extra LLM call per turn (doubles cost, though small)
• What gets extracted depends on the model: occasionally captures one-off details ("user asked about X") → use the 🧠 panel to delete manually
• Embedding coverage: embeddings only cover the memories table (facts); the history table (conversation text) still uses FTS5 keyword search for recall_memory. Full semantic recall over history is a roadmap item.
• No cross-language unification: Chinese extracts Chinese, English extracts English, no translation

Roadmap in IMPROVEMENTS.md.