Kiro Memory
Persistent cross-session memory for AI coding assistants.
Kiro Memory automatically captures what happens during your coding sessions -- files changed, tools used, decisions made -- and injects relevant context when new sessions begin. Your AI assistant remembers what it learned yesterday, last week, and last month.
npm install -g kiro-memory
kiro-memory install
The Problem
Every time you start a new session with an AI coding assistant, it starts from scratch. It does not remember:
- What files you worked on yesterday
- Which architectural decisions were made
- What approaches were tried and rejected
- Where you left off on a multi-session task
Kiro Memory solves this by providing a persistent memory layer that works across sessions and across editors.
Multi-Editor Support
Kiro Memory works with all major AI coding assistants:
| Editor | Integration Method | Auto-Capture | MCP Tools |
|---|---|---|---|
| Claude Code | Hooks + MCP | Yes | Yes |
| Cursor | Rules + MCP | Yes | Yes |
| Windsurf | Rules + MCP | Yes | Yes |
| Cline | Instructions + MCP | Yes | Yes |
| Any MCP editor | MCP server | -- | Yes |
What Your Agent Sees
When a new session starts, Kiro Memory automatically injects relevant context. Here is an example of what your AI assistant receives:
## Kiro Memory — Session Context
### Recent Activity
- [2h ago] Written: src/components/Header.tsx — Added responsive navigation
- [3h ago] Executed: npm run test — 47 passed, 2 failed (auth module)
- [yesterday] Decision: Use React Query for server state management
- [yesterday] Written: src/lib/api.ts — Refactored API client with retry logic
### Active Knowledge
- **Constraint (hard):** No external API calls in SSR components
- **Decision:** Chose Tailwind CSS 4 over styled-components for styling
- **Heuristic:** Run tests before committing changes to auth module
### Last Session Checkpoint
**Task:** Implement user authentication flow
**Progress:** Login/logout working, session persistence pending
**Next steps:** Add refresh token rotation, update middleware
**Files:** src/auth/session.ts, src/middleware/auth.ts
The agent receives this context silently and uses it to provide continuity across sessions.
Key Features
Automatic Context Injection
Context is injected at session start via hooks (Claude Code) or rules files (Cursor, Windsurf, Cline). No manual action required.
Vector Search with Semantic Similarity
Find relevant memories by meaning, not just keywords. Uses local embeddings -- no API keys required.
kiro-memory semantic-search "authentication error handling"
Smart 4-Signal Ranking
Every memory is ranked using four signals to surface the most relevant context:
- Recency -- How recently the observation was created
- Frequency -- How often related files or concepts appear
- Semantic similarity -- How closely the memory matches the current context
- Decay -- Automatic deprecation of stale or outdated observations
Memory Decay
Observations linked to files that have changed since capture are automatically flagged as stale. Consolidation merges duplicate observations, keeping your memory clean and current.
Structured Knowledge
Store architectural decisions, constraints, heuristics, and rejected approaches as first-class knowledge items:
kiro-memory add-knowledge decision "Use PostgreSQL" \
"Chose PostgreSQL over MongoDB for relational data integrity" \
--reason "Need ACID transactions for payment processing" \
--alternatives "MongoDB,SQLite"
Session Checkpoint and Resume
Save progress at any point and resume from where you left off:
kiro-memory resume # Resume latest session
kiro-memory resume 42 # Resume specific session
Activity Reports
Generate weekly or monthly reports of development activity:
kiro-memory report --period=weekly --format=md
Web Dashboard
A real-time analytics dashboard at http://localhost:3001 with:
- Live activity feed via Server-Sent Events
- Spotlight search (Ctrl+K / Cmd+K)
- File hotspot analysis
- Session timeline
- Dark/light theme toggle
10 MCP Tools
A full suite of tools exposed via the Model Context Protocol for programmatic access to search, store, retrieve, and analyze memories.
Full-Text Search
SQLite FTS5-powered keyword search with filters by project, type, and date range.
TypeScript SDK
Programmatic access to the complete Kiro Memory API for building integrations and custom tooling.
Architecture at a Glance
Editor (Claude Code / Cursor / Windsurf / Cline)
|
|-- Hooks / Rules (auto-capture)
|-- MCP Server (10 tools)
|
v
Kiro Memory Worker (port 3001)
|
|-- SQLite + FTS5 (full-text search)
|-- Vector Index (semantic search)
|-- Web Dashboard (real-time UI)
|
v
~/.kiro-memory/kiro-memory.db
Requirements
- Node.js >= 18
- Supported editors: Claude Code, Cursor, Windsurf, Cline, or any MCP-compatible editor