Agentmemory: Persistent Memory Architecture for AI Coding Agents

Working Memory

Every tool invocation gets recorded automatically via PreToolUse and PostToolUse hooks. This includes:

• The command executed
• Input/output content (with privacy filters)
• Execution duration
• Success/failure status

This is the raw data that feeds into higher tiers.

Episodic Memory

After each session ends, working memory gets compressed into an episodic summary. The system uses the configured LLM provider to generate concise session summaries that capture:

• Key files modified
• Problems solved
• Decisions made
• Patterns noticed

Semantic Memory

Over time, episodic memories get analyzed to extract persistent facts:

• Project architecture patterns
• User preferences
• Code style conventions
• Bug fix patterns
• API patterns

This is the “knowledge” that persists across sessions.

Procedural Memory

The highest tier stores workflows — not facts, but how to do things:

• “To add a new component, run X then Y”
• “This project uses convention Z for A”
• “Debug this type of error by checking B first”

Stream 1: BM25

Traditional keyword search with stemming and synonym expansion. Always enabled. This catches exact matches and morphological variants.

Stream 2: Vector Embeddings

Dense embeddings via configurable providers (Voyage, OpenAI, Gemini, local). Cosine similarity search. Automatically enabled when embedding provider is configured.

Stream 3: Knowledge Graph

Entity extraction from queries, then BFS traversal through the relationship graph. Catches semantic connections that keywords miss.

Fusion: Reciprocal Rank Fusion

All three streams feed into RRF with k=60, a parameter tuned for recall. Results are also session-diversified — max 3 results per session to avoid redundancy.

Quick Start

1
# Start the memory server
2
npx @agentmemory/agentmemory

The server exposes:

• REST API on port 3111
• Real-time viewer on port 3113

For OpenCode

Add to opencode.json:

1
{
2
  "mcp": {
3
    "agentmemory": {
4
      "type": "local",
5
      "command": ["npx", "-y", "@agentmemory/mcp"],
6
      "enabled": true
7
    }
8
  }
9
}

For Claude Code

Run in terminal:

1
npx @agentmemory/agentmemory

Then in Claude Code:

1
/plugin marketplace add rohitg00/agentmemory
2
/plugin install agentmemory

For Cursor

Add to ~/.cursor/mcp.json:

1
{
2
  "mcpServers": {
3
    "agentmemory": {
4
      "command": "npx",
5
      "args": ["-y", "@agentmemory/mcp"]
6
    }
7
  }
8
}