Skip to main content
CEMS Banner

​
What is CEMS?

CEMS (Continuous Evolving Memory System) gives AI coding assistants persistent memory across sessions. Instead of starting from scratch every time, your AI assistant remembers your preferences, project conventions, past decisions, and learned patterns.

Quick Start

Get from zero to working CEMS in 5 minutes

Installation

All installation methods and configuration options

Server Deployment

Deploy CEMS server for team usage

API Reference

Complete API and CLI documentation

​
Key Features

Semantic Search

Find memories using natural language, not just keywords. Powered by pgvector and embeddings.

Project-Scoped

Memories automatically boost relevance for the project they were created in.

Multi-IDE Support

Works with Claude Code, Cursor, Codex, Goose, and any MCP-compatible agent.

Auto-Learning

Session end hooks extract learnings automatically. Observer daemon watches transcripts.

Scheduled Maintenance

Nightly consolidation, weekly summarization, monthly re-indexing — all automatic.

Team Memory

Share conventions and decisions with your team using shared memory scope.

​
How It Works

CEMS integrates into your IDE workflow through hooks and MCP tools:

​
Memory Lifecycle

  1. Memory Injection — On every prompt, relevant memories are searched and injected as context
  2. Session Learning — On session end, learnings are extracted and stored
  3. Observational Memory — The observer daemon watches session transcripts and extracts high-level observations
  4. Scheduled Maintenance — Nightly/weekly/monthly jobs deduplicate, compress, and prune memories automatically

​
Architecture Highlights

Everything lives in PostgreSQL with pgvector extension:
  • memory_documents — Documents with user/team scoping, categories, tags
  • memory_chunks — Chunked content with 1536-dim vector embeddings (HNSW index)
  • users / teams — Authentication via bcrypt-hashed API keys
CEMS uses a sophisticated retrieval pipeline:
  1. Query Understanding — LLM routes to vector or hybrid strategy
  2. Query Synthesis — Expands query into 2-5 search terms
  3. HyDE — Generates hypothetical ideal answer
  4. Candidate Retrieval — pgvector HNSW + tsvector BM25
  5. RRF Fusion — Reciprocal Rank Fusion combines results
  6. Relevance Filtering — Removes low-confidence results
  7. Scoring — Time decay, priority boost, project-scoped boost
  8. Assembly — Greedy selection within token budget
Search modes: vector (fast, 0 LLM calls), hybrid (thorough, 3-4 LLM calls), auto (smart routing).
  • 1536 dimensions
  • Batch support for bulk operations
  • Configurable backend (OpenRouter by default)
The cems-observer background process:
  • Polls ~/.claude/projects/*/ JSONL transcript files every 30 seconds
  • Sends 50KB chunks to server for observation extraction
  • Server uses Gemini 2.5 Flash to extract high-level patterns
  • Examples: “User deploys via Coolify”, “Project uses PostgreSQL”

​
Performance

Recall@5: 98%

98% of relevant memories appear in top 5 results

Search Speed

Vector search: <50ms | Hybrid search: <2s

​
Compatibility

Full integration with 6 hooks, 6 skills, 2 commands:
  • Session start: profile + context injection
  • User prompt: memory search + observations
  • Tool use: pre/post hooks for learning and gate rules
  • Session end: learning extraction + observer start
  • Pre-compaction: context preservation

​
Use Cases

1

Personal Preferences

Store coding style, editor settings, workflow preferences:
/remember I prefer Python with type hints and Pydantic for data validation
/remember I use pytest with fixtures for all tests
2

Project Conventions

Document architecture decisions, naming conventions, patterns:
/remember This project uses snake_case for database columns
/remember API endpoints follow REST conventions with /api/v1/...
3

Team Knowledge

Share conventions across the team:
/share We use PostgreSQL for primary database, Redis for caching
/share All PRs require 2 approvals and passing CI
4

Session Context

Automatic learning from your workflow:
  • Successful patterns are extracted at session end
  • Observer daemon captures high-level workflows
  • Maintenance jobs consolidate and compress over time

​
Next Steps

Get Started

Install CEMS client and create your first memory

Deploy Server

Set up CEMS server for your team

Build docs developers (and LLMs) love

Get started for freeTalk to us