contexara 1.0.0

CLI-first memory engine for LLM applications.

Contexara

Memory engine for LLM applications. Drop it into any AI system and it remembers what matters — across sessions, without a vector database, without infra.

pip install contexara
contexara setup   # configure AWS Bedrock credentials

---

How it works

Three tiers of memory, always in sync:

Tier	What it stores	Lifecycle
1 — Raw turns	Every message in the active session	Lives until session closes
2 — Episodes	LLM-crystallized session summaries	Created on session close (background)
3 — Semantic facts	Extracted atomic memories	Permanent, compressed when over budget

Every ask pulls from all three tiers — recent turns for continuity, episode summary for session handoff, semantic memories for persistent facts.

Sessions close after 60 minutes of inactivity. A background worker crystallizes them into structured summaries (title, actions, outcomes, open items) without blocking the user.

Cold archive: raw turns older than 30 days sweep automatically to a SQLite FTS5 store — full-text searchable across years of history.

---

CLI

contexara ask "what was I working on?"
contexara store "prefers Python"
contexara ingest notes.md               # bulk ingest .txt / .md / .jsonl
contexara list
contexara list --since 2026-04-01
contexara search "tech stack"
contexara show <id>
contexara delete <id>
contexara stats
contexara consolidate                   # LLM-compress overlapping memories
contexara prune                         # archive expired (TTL) memories
contexara chat                          # interactive chat mode

Namespaces

Isolate memory per agent, project, or user:

contexara namespace create coding_agent
contexara namespace list
contexara namespace remove coding_agent

contexara ask "what am I building?" --namespace coding_agent
contexara chat --namespace coding_agent

MCP server

Expose Contexara as an MCP tool server — lets AI agents manage their own memory autonomously:

contexara mcp                          # stdio (Claude Desktop, agent SDKs)
contexara mcp --sse --port 8000        # SSE for HTTP agents

13 tools available: chat_ingest, chat_context, memory_store, memory_search, memory_list, memory_forget, memory_deprecate, memory_stats, episodes_search, episodes_latest, session_checkpoint, archive_search, namespace_list.

---

Python SDK

from contexara import ContexaraClient

client = ContexaraClient(namespace="coding_agent")

# Before responding — load context
memories = client.memory.search("user's current goal")
recent = client.chat.get_recent_turns(n=10)
episodes = client.episodes.search("what did I work on last week?")

# After responding — save the turn
client.chat.ingest(user_message, assistant_response)

# When you learn something worth remembering
client.memory.store("user prefers typed Python", kind="preference", importance=3)

# When a major task finishes
client.episodes.crystallize()

All write operations have async equivalents (astore, aingest, acrystallize, etc.).

---

Stack

Layer	Technology
Storage	SQLite — raw turns, episodes, memories, cold archive (FTS5)
Embeddings	Amazon Titan Text Embeddings V2 via Bedrock
LLM	AWS Bedrock — extraction, crystallization, Q&A
Retrieval	RRF hybrid search + temporal weighting
Transports	CLI · Python SDK · MCP (stdio + SSE)

---

Built by Prajwal Narayan