attestari-langchain

LangChain integration for Attestari — the auditable memory layer for AI agents.

Two pieces, both built on stable langchain-core primitives, so instead of replaying a raw chat transcript your chain stores facts with provenance — bi-temporal recall, a tamper-evident audit trail, and a signed deletion certificate, on plain Postgres or zero-dependency in memory:

• AttestariRetriever (BaseRetriever) — recall a subject's relevant facts as Documents, each carrying provenance in its metadata (fact id, source episode, valid-from, confidence). Supports bi-temporal as_of recall. The idiomatic way to inject long-term memory into an LCEL / RAG chain or an agent tool.
• AttestariChatMessageHistory (BaseChatMessageHistory) — a drop-in history for RunnableWithMessageHistory: it learns facts from each human turn (not a raw transcript, and never from the assistant's own words), surfaces the subject's known facts as a system message, and maps clear() to Attestari's provable deletion (forget).

Install

pip install attestari-langchain        # pulls in attestari + langchain-core

Use — inject memory into a chain (AttestariRetriever)

from attestari import Memory
from attestari_langchain import AttestariRetriever

mem = Memory()                       # or Memory.local() / Memory.postgres()
mem.add("Hi, I'm Alice. I live in Toronto.", subject_id="user_42")

retriever = AttestariRetriever(
    mem=mem,                         # the Attestari engine
    subject_id="user_42",            # an opaque pseudonym, not raw PII
    k=5,                             # facts to recall
    # as_of="2022-01-01",            # optional: recall what was true then
)

docs = retriever.invoke("where does the user live?")
# -> [Document("user_42 lives in Toronto", metadata={fact_id, source_episode_id, score, ...})]

The zero-dependency default extractor understands first-person statements ("I live in…", "my name is…"). For arbitrary third-person text, install attestari[anthropic] and set ANTHROPIC_API_KEY for Claude-powered extraction.

Drop retriever into any LCEL chain the way you would any other retriever; each returned Document carries provenance in .metadata you can show or audit.

Use — conversational memory (AttestariChatMessageHistory)

from attestari import Memory
from attestari_langchain import AttestariChatMessageHistory
from langchain_core.messages import HumanMessage

history = AttestariChatMessageHistory(mem=Memory(), subject_id="user_42")
history.add_messages([HumanMessage("Hi, I'm Alice and I live in Toronto.")])
history.messages          # -> [SystemMessage("Known facts about the user:\n- user lives_in Toronto")]
history.clear()           # provable deletion (forget) for this subject

Wire it into RunnableWithMessageHistory as the get_session_history factory:

from langchain_core.runnables.history import RunnableWithMessageHistory

chain_with_memory = RunnableWithMessageHistory(
    chain,
    lambda session_id: AttestariChatMessageHistory(mem=Memory.local(), subject_id=session_id),
    input_messages_key="input",
    history_messages_key="history",
)

Constructor reference

AttestariRetriever

AttestariChatMessageHistory(mem, subject_id) — add_messages(msgs) ingests each human turn (assistant messages are conversation, not testimony — they are not attributed to the user as facts), messages returns the subject's live facts as one system message, and clear() maps to provable deletion.

A runnable, no-LLM example is in example.py: python clients/langchain/example.py.

Apache-2.0.