contexara 2.2.0

AI that forgets is just autocomplete. Contexara gives your agent a memory it can build on.

Contexara

AI that forgets is just autocomplete. Contexara gives your agent a memory it can build on.

Persistent, three-tier memory engine for AI agents. Drop it in and your agent remembers facts, corrections, and past sessions — automatically, across every conversation.

Install

pip install contexara
contexara setup        # configure AWS Bedrock credentials
contexara ask "what was I working on?"

Run setup on first install:

contexara setup

This guides you through AWS Bedrock credentials and saves them to ~/.contexara/.env.

How it works

Tier	What it stores	When
L0 — Raw turns	Every user/assistant exchange verbatim	Every turn
L1 — Episodes	LLM-crystallized session summaries	After session ends
L2 — Semantic memory	Typed facts: preferences, corrections, constraints, tech choices	Every turn

On session start, all three tiers inject context automatically — your agent is never blind.

Python SDK

from contexara import ContextaraClient

mem = ContextaraClient(namespace="my-agent")

ctx = mem.context()                        # inject history + last episode
facts = mem.retrieve("user preferences")  # hybrid search before answering
mem.ingest(user_text, assistant_text)     # extract + store after every turn
mem.store("User prefers bullet points", kind="style")  # explicit store
mem.checkpoint()                          # crystallize session now

MCP Server

Works with Claude Desktop and any MCP-compatible framework:

{ "mcpServers": { "contexara": { "command": "python", "args": ["-m", "contexara.mcp_server"] } } }

What's in v2.1.0

• 9 memory kinds — profile, correction, preference, constraint, tech_preference, style, pattern, task, note
• Zero fact-loss crystallization — 3-pass LLM pipeline with fact audit and repair
• Hybrid search — FTS5 + cosine vector RRF on hot DB, FTS5 on cold archive
• Never-delete versioning — full history of every memory update
• Namespace isolation — separate memory space per agent or application
• Model routing — fast model for chat, judge model for extraction and evals
• ~/.contexara/ storage — persists across projects

---

Memory model

Tier	Stores	When
Raw turns	Every message in the session	Live
Episodes	LLM-crystallized session summaries	On session close
Semantic facts	Extracted atomic memories	Permanent
Cold archive	Turns older than 30 days	FTS5, auto-swept

Sessions idle for 60 minutes crystallize automatically in the background.

---

Interfaces

CLI — ask, store, search, list, chat, ingest, namespace, mcp, eval

MCP server — 13 tools over stdio or SSE. Plug into Claude Desktop or any agent SDK.

Python SDK

from contexara import ContexaraClient
client = ContexaraClient(namespace="my_agent")
client.memory.store("user prefers Python", kind="preference")
memories = client.memory.search("current goals")
client.chat.ingest(user_msg, assistant_msg)

Dashboard

contexara dashboard --port 8000

Latency charts, trace explorer, namespace management, LLM-as-judge eval scores — all in one UI.

---

Stack: SQLite · AWS Bedrock · Amazon Titan Embeddings · RRF hybrid retrieval · FastAPI · React

Built by Prajwal Narayan