tessera-memory 0.1.4

Long-term memory for LLM agents: the typed, sync and async Python client for the Tessera memory service.

tessera-memory

Long-term memory for LLM agents.

Your agent forgets everything between sessions. Tessera remembers. Write conversational turns, then recall ranked hits or a prompt-ready context block. Sync and async, fully typed.

Getting started · Concepts · API reference · Use with MCP

What you get

🧠 Memory	Extracts facts and episodes from conversation turns. Handles corrections, pinning, and forgetting.
🔍 Search	Ranked, typed retrieval over your memory. No LLM in the loop, so it stays cheap and deterministic.
💬 Query	Composes retrieved memory into a prompt-ready context block for your model.
📚 Procedures	Store reusable lessons (trigger, steps, outcome) and recall them by task.
⚡ Sync + async	The same typed API on Tessera and AsyncTessera.
🔌 MCP built in	Give Claude Code, Codex, and Cursor memory with one server.

Install

pip install tessera-memory   # or: uv add tessera-memory

Requires Python 3.10+.

Quickstart

from tessera_memory import Tessera

client = Tessera()  # reads TESSERA_API_KEY, or Tessera(api_key="tsk_live_...")

# Write what happened.
client.memories.add(content="Ada prefers dark roast coffee.", role="user", user_id="ada")

# Get ranked hits (no LLM)...
hits = client.search(query="what coffee does Ada like?", user_id="ada")

# ...or a prompt-ready context block (may call an LLM server-side).
answer = client.query(query="what coffee does Ada like?", mode="chat", user_id="ada")
print(answer.context)

That is the whole loop: add to remember, search to retrieve, query to get context for a model.

Async

Every method exists on AsyncTessera with await:

import asyncio
from tessera_memory import AsyncTessera


async def main() -> None:
    async with AsyncTessera() as client:
        await client.memories.add(content="Ada prefers dark roast coffee.", role="user", user_id="ada")
        async for item in client.memories.list(user_id="ada"):
            print(item.text)


asyncio.run(main())

Use with MCP (Claude Code, Codex, Cursor)

Give your coding agent memory with the tessera-mcp server. uvx fetches it (and the SDK) on first run — your repo needs no Python, pip, or virtualenv; it all runs in an isolated uvx environment. The only machine-wide prerequisite is uv (a single binary that downloads its own Python if needed). Two more things are handled for you:

• Repo isolation is automatic. Memory is namespaced per git repo (from the origin remote, falling back to the folder name). You never declare a repo name.
• Your key is stored once, in a file. Run tessera-mcp login and the key is saved to ~/.tessera/credentials.json — no shell export, no ${VAR} in config, no restart.

Claude Code (one-step plugin). Install the plugin, then save your key once:

/plugin marketplace add harshkedia177/tessera-python
/plugin install tessera-memory@tessera

uvx --from tessera-mcp tessera-mcp login   # paste your tsk_live_... key when prompted

That's it — you get the five memory_* tools, auto-recall hooks, and the skill. (If you skip the login step, the first memory call simply tells the agent to ask you for the key and run it for you.)

Cursor / Claude Desktop. Add the server (no env block needed), then run the same tessera-mcp login once:

{
  "mcpServers": {
    "tessera": {
      "command": "uvx",
      "args": ["--from", "tessera-mcp", "tessera-mcp"]
    }
  }
}

Codex. Add to ~/.codex/config.toml, then run tessera-mcp login once:

[mcp_servers.tessera_memory]
command = "uvx"
args = ["--from", "tessera-mcp", "tessera-mcp"]

Prefer the environment? TESSERA_API_KEY (key) and TESSERA_REPO (to override the auto-detected namespace) still work everywhere and take precedence over the stored values.

Full details — the all-in-one Claude Code plugin, hooks, and the privacy notes on transcript consolidation — are in Use with MCP.

Configuration

The client reads TESSERA_API_KEY from the environment, or you pass it directly. Auth goes out as a bearer token. For timeouts, retries, logging, raw responses, and a custom HTTP client, see Configuration.

Documentation

• Getting started: install, connect, first reads and writes.
• Concepts: turns, episodes, facts, procedures, resources.
• Configuration: clients, retries, timeouts, logging, transport.
• Guides: memories · search and query · procedures and resources · error handling
• API reference: the full method surface.

Development

make sync-spec    # copy openapi.json from the server checkout
make generate     # regenerate src/tessera_memory/models.py from openapi.json
make lint         # ruff check + format --check
make typecheck    # mypy --strict
make test         # pytest (respx-mocked)

License

Apache 2.0. See LICENSE.