contexara 2.4.0

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 layer for AI agents. Drop it in and your agent remembers facts, preferences, and past sessions — automatically, across every conversation.

---

Install

pip install contexara
contexara setup
contexara ask "what was I working on?"

For specific providers:

pip install contexara[openai]
pip install contexara[anthropic]
pip install contexara[gemini]
pip install contexara[all-providers]

---

How It Works

Tier	What	Scope
T1 — Raw Turns	Every user/assistant message, stored verbatim	Current session
T2 — Episodes	LLM-crystallized summaries of past sessions	Cross-session
T3 — Semantic Memory	Extracted facts, preferences, constraints, corrections	Permanent

On every query, all three tiers inject context automatically — your agent is never blind.

---

Supported Providers

Provider	CONTEXARA_PROVIDER value
AWS Bedrock	bedrock (default)
OpenAI	openai
Anthropic	anthropic
Google Gemini	gemini
OpenRouter	openrouter

Configure via contexara setup or manually in ~/.contexara/.env:

CONTEXARA_PROVIDER=openai
CONTEXARA_MODEL=gpt-4o
OPENAI_API_KEY=sk-...

---

Python SDK

from contexara import ContextaraClient
from contexara.llm import call_llm

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

while True:
    user_input = input("You: ").strip()
    if user_input.lower() in ("exit", "quit"):
        mem.flush()   # extract memories from remaining turns on exit
        break

    prompt = mem.build_prompt(user_input)   # retrieve + enhance (T1+T2+T3)
    reply = call_llm(prompt)
    print(f"Assistant: {reply}")

    mem.ingest(user_input, reply)           # store turn + batch extract every 10 turns

Additional SDK Methods

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

# Explicit memory operations
mem.store("User prefers bullet-point responses", kind="preference")
mem.retrieve("what does the user prefer?")   # returns [{content, kind}, ...]
mem.search("budget constraints")             # returns full records with metadata
mem.stats()                                  # memory counts by kind and source
mem.checkpoint()                             # force crystallize current session
mem.history(memory_id)                       # version history for a memory

---

CLI Reference

Command	Description
contexara setup	Configure AI provider credentials
contexara ask "<query>"	Single query with full memory context
contexara chat	Interactive chat mode
contexara store "<text>"	Save a memory directly
contexara list	List stored memories
contexara search "<query>"	Semantic search over memories
contexara stats	Memory counts by kind and source
contexara consolidate	LLM-merge overlapping memories
contexara clear	Delete all memories in namespace
contexara namespace list|create|switch|delete	Manage namespaces
contexara serve	Start REST API server
contexara dashboard	Start web dashboard
contexara mcp	Start MCP server

---

Namespaces

Complete data isolation between agents or projects:

contexara namespace create work
contexara namespace switch work
contexara ask "what are my pending tasks?" --namespace work

---

Key Features

• Local first — all data in ~/.contexara/, no cloud required
• LLM agnostic — Bedrock, OpenAI, Anthropic, Gemini, OpenRouter
• Batch extraction — memories extracted every 10 turns for cost efficiency, flushed on exit
• Intent classification — detects historical vs. recent vs. semantic queries, adjusts retrieval strategy
• Hybrid retrieval — FTS5 + cosine vector search with RRF fusion
• Multiplicative scoring — memory ranked by similarity × importance × recency, per-kind decay
• Memory observability — tracks which memories were cited, usage ratio, and confidence signals per turn
• Session continuity — sessions crystallize after 60 min inactivity, injected on next session start
• Never-delete versioning — full history of every memory update
• Cold archive — turns older than 30 days auto-swept to FTS5-searchable archive
• MCP server — plug into Claude Desktop or any MCP-compatible framework
• Web dashboard — memory browser, session explorer, latency charts, LLM-as-judge evals

---

Data Storage

~/.contexara/
├── memory.db          # T3 semantic memories
├── active_state.db    # T1 raw turns + T2 episodes + sessions
└── cold_archive.db    # archived turns older than 30 days

---

Built by Prajwal Narayan