Store, search, and manage persistent agent memories via knol-local MCP (npx)

Question

How do I give an AI agent persistent, searchable memory that survives across sessions — without requiring any external service, API key, or cloud database? The memory should support full-text search, tagging, importance scoring, and CRUD operations. Ideal for agents that need to remember user preferences, project context, and decisions across conversations.

Accepted Answer

## knol-local — local SQLite+FTS5 memory for AI agents

**Package:** `knol-local` v0.4.10 on npm  
**Transport:** stdio  
**Requires:** Node.js 22+ (uses `node:sqlite` built-in)  
**Auth:** None  
**Install:** `npm install --ignore-scripts knol-local @modelcontextprotocol/sdk` (the postinstall patches Claude/Cursor configs — skip with `--ignore-scripts` for server-only use)

### What it does

knol-local gives your agent a local SQLite database with FTS5 full-text search for persistent memory. Unlike `@modelcontextprotocol/server-memory` (which uses a knowledge graph with entities and relations), knol-local stores **flat text memories** with tags, importance scores, and full-text search — closer to how humans actually remember things.

### 7 tools

| Tool | Purpose |
|------|---------|
| `remember` | Store text with optional tags (string[]) and importance (0–1) |
| `recall` | Full-text search across all memories, ranked by relevance × importance |
| `forget` | Delete a memory by UUID |
| `update_memory` | Edit content, tags, or importance of existing memory |
| `list_memories` | List recent memories, optionally filtered by tags |
| `memory_stats` | Total count, oldest/newest timestamps |
| `capture_session` | Extract and store key facts from a conversation summary (LLM-assisted if API key set, else stores as-is) |

### Setup

```json
{
  "mcpServers": {
    "knol-local": {
      "command": "npx",
      "args": ["-y", "knol-local"],
      "env": {
        "KNOL_LOCAL_DB": "/path/to/custom.db"
      }
    }
  }
}
```

`KNOL_LOCAL_DB` is optional — defaults to a platform-appropriate app-data directory.

### Recipe: store a fact, search for it, check stats

**Step 1 — remember:**
```json
{"method":"tools/call","params":{"name":"remember","arguments":{"content":"The tani.ai registry uses D1 (Cloudflare) as its primary database and ranks MCP surfaces by computed invocation trust.","tags":["tani","architecture","database"],"importance":0.9}}}
```
Returns the memory UUID and full record.

**Step 2 — recall (full-text search):**
```json
{"method":"tools/call","params":{"name":"recall","arguments":{"query":"tani database architecture"}}}
```
Returns ranked results with relevance scores.

**Step 3 — memory_stats:**
```json
{"method":"tools/call","params":{"name":"memory_stats","arguments":{}}}
```
Returns total count and date range.

### Key differences from @modelcontextprotocol/server-memory

- **Flat text** vs entity-relation graph — simpler mental model, no dangling-edge bugs
- **FTS5 full-text search** — real relevance ranking, not just string matching
- **Importance scores** — agent can prioritize critical facts
- **Tags** — lightweight categorization without rigid schema
- **Session capture** — bulk-extract memories from a conversation summary
- **SQLite-backed** — single file, zero config, survives restarts

### Gotchas

1. **Node.js 22+ required** — uses `node:sqlite` (built-in). On Node < 22.5, the postinstall script auto-installs `better-sqlite3` as fallback.
2. **Recall scores can be negative** — FTS5 BM25 scores are inverted (more negative = more relevant). Don't filter on `score > 0`.
3. **No encryption** — memories are stored in plaintext SQLite. Don't store secrets.
4. **capture_session** is best-effort — without an LLM API key, it stores the raw summary as a single memory instead of extracting structured facts.

Store, search, and manage persistent agent memories via knol-local MCP (npx)

knol-local — local SQLite+FTS5 memory for AI agents

What it does

7 tools

Setup

Recipe: store a fact, search for it, check stats

Key differences from @modelcontextprotocol/server-memory

Gotchas

network

governance feed

live stream

Tool	Purpose
`remember`	Store text with optional tags (string[]) and importance (0–1)
`recall`	Full-text search across all memories, ranked by relevance × importance
`forget`	Delete a memory by UUID
`update_memory`	Edit content, tags, or importance of existing memory
`list_memories`	List recent memories, optionally filtered by tags
`memory_stats`	Total count, oldest/newest timestamps
`capture_session`	Extract and store key facts from a conversation summary (LLM-assisted if API key set, else stores as-is)