Private Beta

Memory for AI Agents

Persistent context that compounds across sessions. Your agents remember patterns, learn from mistakes, and get smarter over time.

Terminal
5
Search Channels
50+
MCP Tools
40/s
Embeddings
<3s
Search Time
How it works

Four commands to memory

From initialization to production in minutes. No complex setup, no configuration files, no infrastructure to manage.

01

Initialize

cas init

Set up CAS in any project. Zero config required. Creates a local SQLite store.

02

Remember

cas remember "..." -t fact

Store entries with epistemic types. Facts, opinions, hypotheses — know what you know.

03

Retrieve

cas search "query"

5-channel hybrid search. BM25, semantic, temporal, graph, and reranking.

04

Serve

cas serve --mcp

Expose 50+ MCP tools. Works with Claude Code, Cursor, and any MCP client.

Architecture

Built for agent cognition

Not just storage. A cognitive architecture inspired by MemGPT and Hindsight, designed for how AI agents actually work.

Memory Tiering

InContext → Working → Cold → Archive. Memories flow based on access patterns and the forgetting curve.

5-Channel Search

BM25 + Semantic + Temporal + Graph + Reranking. Each channel weighted by query context.

GPU Accelerated

Qwen3 embeddings on Metal. 40 embeddings per second locally, zero API costs.

Epistemic Types

Facts, opinions, hypotheses. Your agent knows what it knows and what it doesn't.

Rule Lifecycle

Draft → Proven → Stale → Retired. Auto-sync rules to Claude Code CLAUDE.md.

MCP Native

50+ tools exposed via MCP. Works with Claude Code, Cursor, and any compatible client.

The Forgetting Curve

Memories decay like human memory

Each entry has retrievability that decays over time unless reinforced. Access patterns determine tier promotion. The system learns what matters.

Live Decay Model
1.00.50R=0.3Time →
R = e^(-t/S)
Retrievability
S = f(reps, lapses)
Stability
R < 0.3 → demote
Threshold
cas search --explain
# Hybrid search with channel weights
$ cas search "memory safety" --explain
Query type:technical, recent context
Channel weights:
BM25
Semantic
Temporal
Graph
Results:
"Use memory-safe abstractions"0.94
"Avoid raw pointers in handlers"0.87
"Consider arena allocation"0.71
Retrieved 3 entries in 47ms

Stop teaching agents
the same things twice

Join the waitlist for early access. We're onboarding teams weekly.

1
2
3
Open source
MIT licensed
Self-hostable
Your infrastructure
Local-first
Offline capable