Tencent Agent Memory

TencentDB Agent Memory

Source: Tencent/TencentDB-Agent-Memory (Apache-2.0)

Tier: TIER 2 — CORRECTNESS

Bộ nhớ dài hạn 4 tầng cho AI agent — không cần API ngoài, token giảm 61%, success rate tăng 51%.

Do NOT use for: hermes-memory-manager (multi-provider orchestration), mem0 (managed cloud memory), terminal--agent-memory (general patterns).

Kiến trúc 4 tầng (L0 → L3)

L0 — Raw conversations       Raw tool call logs + full conversation history
       ↓ compress
L1 — Atomic facts            Extracted facts, decisions, preferences (SQLite)
       ↓ cluster
L2 — Scenarios/scenes        Grouped contexts — "debugging session", "PR review"
       ↓ abstract
L3 — Personas                High-level user/project profile updated over time

Retrieval: L3 (fast, broad) → L2 (context) → L1 (precise) → L0 (evidence)

Mỗi tầng human-readable → dễ debug. Đường dẫn từ L3 abstraction xuống L0 raw là deterministic — không bao giờ mất trace.

Short-term memory: Mermaid tool-log compression

Tool logs dài và tốn token. TencentDB nén chúng thành Mermaid state diagram:

<!-- Thay vì lưu 200 dòng tool output: -->

```mermaid
stateDiagram-v2
    [*] --> ReadFile: read package.json
    ReadFile --> RunTests: npm test
    RunTests --> FAIL: 3 failures
    FAIL --> EditFile: fix auth.ts:42
    EditFile --> RunTests: re-run
    RunTests --> PASS: all green
    PASS --> Commit: git commit

Token savings: ~60% vs raw logs, vẫn giữ đủ context để tiếp tục task.

function compressToolLog(calls: ToolCall[]): string {
  const lines = calls.map((c, i) => {
    const from = i === 0 ? '[*]' : calls[i-1].tool
    const status = c.error ? 'FAIL' : (c.result?.slice(0, 20) ?? 'ok')
    return `    ${from} --> ${c.tool}: ${status}`
  })
  return `\`\`\`mermaid\nstateDiagram-v2\n${lines.join('\n')}\n\`\`\``
}

Long-term memory: SQLite + sqlite-vec

// Setup — local only, no API key
import Database from 'better-sqlite3'
import { load as loadVec } from 'sqlite-vec'

const db = new Database('.yana/agent-memory.db')
loadVec(db)

db.exec(`
  CREATE VIRTUAL TABLE IF NOT EXISTS facts USING vec0(
    embedding float[768]
  );
  CREATE TABLE IF NOT EXISTS facts_meta (
    id INTEGER PRIMARY KEY,
    tier INTEGER,       -- 1=atomic 2=scenario 3=persona
    content TEXT,
    tags TEXT,
    ts INTEGER
  );
`)

// Upsert atomic fact (L1)
function storeFact(content: string, embedding: number[], tags: string[]) {
  const meta = db.prepare(
    'INSERT INTO facts_meta(tier,content,tags,ts) VALUES(1,?,?,?)'
  ).run(content, tags.join(','), Date.now())
  db.prepare('INSERT INTO facts(rowid,embedding) VALUES(?,?)').run(
    meta.lastInsertRowid, new Float32Array(embedding)
  )
}

// Hybrid search: BM25 + vector với RRF fusion
function recall(query: string, embedding: number[], k = 5) {
  const vec = db.prepare(`
    SELECT rowid, distance FROM facts
    WHERE embedding MATCH ? ORDER BY distance LIMIT ?
  `).all(new Float32Array(embedding), k * 2)

  const bm25 = db.prepare(`
    SELECT id FROM facts_meta WHERE content LIKE ? LIMIT ?
  `).all(`%${query}%`, k * 2)

  // RRF fusion: score = Σ 1/(rank + 60)
  const scores = new Map<number, number>()
  vec.forEach(({ rowid }, i)  => scores.set(rowid, (scores.get(rowid) ?? 0) + 1/(i+61)))
  bm25.forEach(({ id }, i)    => scores.set(id,    (scores.get(id)    ?? 0) + 1/(i+61)))

  return [...scores.entries()]
    .sort((a, b) => b[1] - a[1])
    .slice(0, k)
    .map(([id]) => db.prepare('SELECT * FROM facts_meta WHERE id=?').get(id))
}

Persona builder (L3)

// L3 persona — built từ accumulated L1/L2 facts
interface AgentPersona {
  projectContext: string    // "yana-ai: Rust runtime + 3518 skills, multi-agent"
  workStyle:      string    // "prefers surgical edits, no comments, vi + en"
  recentFocus:    string[]  // ["mobile sync", "security rules", "README cleanup"]
  avoidPatterns:  string[]  // ["don't say oke ngon lành while bugs remain"]
}

// Inject persona vào system prompt — thay thế manual context trong mỗi session
function buildSystemContext(persona: AgentPersona): string {
  return [
    `Project: ${persona.projectContext}`,
    `Style: ${persona.workStyle}`,
    `Recent: ${persona.recentFocus.join(', ')}`,
    `Avoid: ${persona.avoidPatterns.join('; ')}`,
  ].join('\n')
}

Tích hợp với Yana AI L1/L2

TencentDB L0–L3 mở rộng hệ thống Yana AI hiện có:

Triển khai: L0/L1 dùng Yana AI cũ, thêm L2 clustering + L3 persona update cuối session.

Cài đặt

npm install better-sqlite3 sqlite-vec

Anti-Fake-Pass Checks

❌ FAIL nếu dùng external API cho embedding trong local-only deployment
❌ FAIL nếu Mermaid output không thể parse lại thành tool call sequence
❌ FAIL nếu L3 persona overwrite L1 facts (phải keep cả hai)
❌ FAIL nếu RRF fusion bỏ qua một trong hai signals (BM25 hoặc vector)
✅ PASS khi: recall() trả về kết quả trong < 50ms trên 10K facts
✅ PASS khi: Mermaid diagram < 30% kích thước raw tool log gốc

See also

• hermes-memory-manager — multi-provider orchestration (không cover L0–L3 pipeline)
• mem0 — managed cloud memory (cần API, khác local SQLite approach)
• mermaid-diagram-generation — Mermaid cho docs (không phải tool log compression)
• memory-persistence-law.md — Yana AI rule về L1/L2 persistence