AIrecall

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

AIrecall

A drop-in long-term memory layer for AI agents. Episodic + semantic memory, hybrid retrieval, and auto-summarization — so your agent remembers what matters across sessions.

pip install airecall-sdk
airecall init

Languages	Python SDK + Go memory server
Storage	SQLite + in-process vector index
License	MIT
Dependencies	SDK stdlib-only · server one pure-Go dep

---

Why This Exists

Most agent frameworks give you a context window, not a memory. Close the session and everything the agent learned — user preferences, past decisions, corrections it was given — is gone. Stuff it all into the prompt instead, and you’re paying for and diluting your context with old information, most of which is irrelevant to the current turn.

AIrecall sits between your agent and a persistent store, and gives it an actual memory:

🧠 Episodic memory	What happened, in order (conversations, actions, outcomes)
📌 Semantic memory	Durable facts and preferences distilled out of those episodes
🔎 Hybrid retrieval	Keyword + vector search pulls back what matters for this turn
🗜️ Auto-summarization	Old episodes are compressed, not deleted — long-term memory stays cheap

The goal is a memory layer that’s boring to integrate and hard to notice — until you turn it off and the agent forgets your name.

---

How It Works

        +---------------------+        local call / gRPC        +----------------------+
        |      Agent code      | -------------------------------> |     AIrecall Core     |
        |    (Python SDK)      | <------------------------------- |   (memory server)     |
        +---------------------+                                  +-----------+----------+
                                                                             |
                                                                 +-----------v----------+
                                                                 |    Storage Engine      |
                                                                 |  SQLite + vector idx   |
                                                                 +-----------+----------+
                                                                             |
                                                                 +-----------v----------+
                                                                 |   Optional MCP         |
                                                                 |   adapter              |
                                                                 +-----------------------+

On every turn, the SDK sends the current query to the core, which does a hybrid retrieval pass — keyword + vector similarity — over both episodic and semantic memory, and returns the top-k relevant memories to inject into your prompt. In the background, a summarizer periodically walks older episodes, extracts durable facts into semantic memory, and compacts the rest.

---

Install

# Python SDK - what your agent code imports
pip install airecall-sdk

# Core memory server - runs locally or as a sidecar
pip install airecall-sdk[server]
# or run it standalone:
airecall serve

airecall init scaffolds a local SQLite-backed store so you can start storing and recalling memories immediately, no separate service required.

---

Quickstart

from airecall import Memory

memory = Memory(agent_id="support-bot")

# Store an episode
memory.remember(
    "User asked about refund policy for order ORD-9921, told 30-day window applies."
)

# Later, in a new session
context = memory.recall("what did we tell this user about refunds?")
# -> returns the relevant episodic memory, ranked by relevance

# Promote a durable fact explicitly
memory.remember_fact(key="preferred_contact", value="email, not phone")

Framework adapters:

from airecall.adapters.langchain import MemoryRetriever

retriever = MemoryRetriever(memory)

## Troubleshooting

- **`RecallTimeout` on first query** - the index is still warming. Retry after the health endpoint reports `status: ready`.
- **Missing memories after restart** - check the `storage.backend` path; a relative path resolves against the working directory of the server process.