polign-recall

A Python client for Recall, correctable agent memory. Requires Python 3.10+.

pip install polign-recall

That is the whole install. pip also brings the polign_db package, which holds the polign CLI and polign-server binaries for Linux, macOS and Windows, so there is nothing else to download.

from polign_recall import Client

with Client(local_dir="./recall-data", extract_model="anthropic:claude-haiku-4-5-20251001") as memory:
    memory.remember(text="I switched from helix to zed last week.")
    for m in memory.ask("Which editor do I use?"):
        print(m.text, m.since, [e.text for e in m.before])  # user prefers editor: zed ... ['helix']
    memory.forget_text("my editor")

remember(text=...) hands the server a passage; the model named by extract_model works out the facts in it, and a newer fact replaces an older one. It returns a TextResult listing what was filed and what each statement replaced. ask returns Memory sentences, each with since, days_ago, and what it replaced under before; it accepts as_of. forget_text withdraws the facts a description names and returns them; the texts they came from stay as the record. Without extract_model, remembered text is kept as written: ask finds it by its words, but it replaces nothing, and forget_text needs a model.

Nothing is deleted, so the record of corrections can be read back:

for t in memory.history_text("my editor"):       # every statement, oldest first
    for step in t.changes:
        print(step.when, step.change, step.ended, step.began)
memory.ask("Which editor do I use?", as_of="2026-10-01T00:00:00Z")  # believed then
memory.diff(since="2026-10-01T00:00:00Z")           # what ended and began since

history_text returns a Timeline per thing the question reaches, including things since forgotten: now is what is believed, and each Step says whether it stated, replaced or withdrawn something, with the sentences it ended and began. diff(since, until=None) returns a Change per thing whose beliefs differ between the two instants. Both need a recall server newer than 0.21; an older one raises RecallError with code unsupported.

local_dir keeps the database on this machine. The first client to open the directory starts a polign-server for it in the background, listening on localhost only and protected by a key stored in the directory. Later clients, including ones in other processes, share that server. It keeps running after your program exits; its process id is in runtime.json and its log in server.log, both inside the directory. Managed local databases work on Linux and macOS.

backend, backend_url and backend_key name any database, the same three names as recall -backend, -url and -key: backend is "local", "polign", "pgvector", "mongodb" or "qdrant", the URL is in that backend's own form, and the key is the credential when the URL does not carry one. local_dir=X is the short spelling of backend="local", backend_url=X.

Client(backend="polign", backend_url="http://memory.internal:23000", backend_key="plgn_...")
Client(backend="pgvector", backend_url="postgres://recall:pw@db.internal/recall?schema=memory")
Client(backend="qdrant", backend_url="https://x.cloud.qdrant.io:6333", backend_key="...")

Given any of the three, the client hands the binary exactly those and nothing the environment says about a database. Given none, the recall binary resolves it: RECALL_BACKEND, RECALL_BACKEND_URL and RECALL_BACKEND_KEY in the environment, then what recall setup saved, then a local database in ~/.local/share/recall that it runs itself and that every agent on the machine shares. So a bare Client() works on a fresh machine with nothing configured. The names used before recall 0.23 (POLIGN_URL and the rest) still work in the environment, and a client given backend= against an older binary says the same thing in the names that binary knows. Client.database says where memory ended up, as the binary resolved it: a Database with the backend, the url (a directory for local) and, for local, the loopback server address; None from a binary older than 0.23 or a recall serve.

The client owns a long-lived recall mcp -write -text subprocess and shares the Go implementation's validation and fold. Use write=False for a read-only connection. It runs the recall binary that this package's platform wheel installs next to python; on a platform without one it runs recall from PATH, and without recall it falls back to polign mcp -memory-only from polign_db 0.13 or earlier. command=[...] overrides all of these.

The model is provider:model: anthropic:<model>, openai:<model>, or ollama:<model>, with keys from ANTHROPIC_API_KEY or OPENAI_API_KEY. POLIGN_EXTRACT_MODEL in the environment does the same.

Typed memory

With typed=True, the client uses the typed tools (recall mcp -typed, recall 0.18 or later): your code names the subject, predicate and value, over a fixed vocabulary.

with Client(local_dir="./recall-data", typed=True) as memory:
    memory.remember("user", "prefers_editor", "vim")
    changed = memory.remember("user", "prefers_editor", "neovim")
    current = memory.recall("user", "prefers_editor")[0]
    print(current.value, current.replaced[0].value)  # neovim vim
    print(memory.history("user", "prefers_editor"))  # both statements
    for step in memory.timeline("user", "prefers_editor").changes:
        print(step.when, step.change, step.ended, step.began)  # ... replaced vim -> neovim
    memory.forget("user", "prefers_editor", "neovim")

history returns the raw events; timeline tells the same record as changes, in the same Timeline and Step types history_text returns, with subject and predicate set. diff(since, until=None, subject=None) works here too, and each Change also carries its subject, predicate, and the beliefs before and after.

Read predicates() for the vocabulary: the starter registry, or the file named by POLIGN_PREDICATES. remember(text=..., statements=[...]) accepts proposals from your agent's model; every statement must have subject, predicate, typed value, and evidence quoting the text exactly. A proposal that names an unlisted predicate is dropped, and the whole text is kept as a note either way. With extract_model and no statements, the server's model proposes them over the same vocabulary.

The text calls (ask, forget_text, remember(text=...) alone) raise ValueError on a typed client, and the typed calls (recall, forget, history, predicates) raise on a text client. With command=[...] or url=..., the client follows whichever tools the server lists; typed=True or text=True against the other surface is an error at open.

recall(..., as_of=...) accepts an RFC3339 string or timezone-aware datetime. forget requires a typed value or explicit all=True; false and zero remain values. Forgetting retracts a belief and retains its historical events.

Errors raise RecallError with a code and, for a failed extraction batch, any reported partial results. Calls are serialized, timeout after 120 seconds by default, and never automatically retry writes. A timeout closes the transport; inspect history using a new client before retrying an uncertain write.

Connecting to a server

When memory is served by recall serve rather than a local process, give the client the server and one credential and it speaks HTTP instead:

from polign_recall import Client

client = Client(url="http://recall.recall.svc:8080", api_key="rcl_...")

Inside a Kubernetes sandbox the credential is the pod's projected service account token, which the server exchanges for a short-lived key and the client renews:

client = Client(url="http://recall.recall.svc:8080", token_file="/var/run/secrets/recall/token")

Client.from_env() reads RECALL_URL, RECALL_API_KEY and RECALL_TOKEN_FILE (default /var/run/secrets/recall/token when it exists) and falls back to a local process when RECALL_URL is unset, so the same program runs on a laptop and in a sandbox. Over HTTP the server decides whether it serves the text or the typed tools and whether the resume tools exist; the client finds out when it connects.

Resuming an agent

An agent that crashes or moves to another machine can pick up from its own records instead of a snapshot of its process. Open the client with agent=True and call resume:

from polign_recall import Client

with Client(local_dir="./recall-data", agent=True) as client:
    with client.resume("billing-migrator") as agent:
        print(agent.context.briefing)  # hand this to the model as its starting context
        agent.update_working_state(goal="Move billing to the v2 API",
                                   plan=["find call sites", "migrate them", "run the tests"])
        agent.record_turn("user", "Start with the charge endpoint.")
        agent.set_pointer("wip", "git_ref", {"repo": "github.com/acme/billing", "branch": "agent/wip"})
        agent.milestone("call sites found", progress="12 call sites listed")

The first resume of an id starts the agent fresh (context.fresh is true). Every later one rebuilds context from what the agent wrote: its working state, pointers to its work, relevant memories, and its most recent turns, trimmed to token_budget. context.briefing is all of that as one text for the model; the same parts are also there as fields.

resume takes a lease on the agent id, so only one process acts for it at a time. While another process holds it, resume raises RecallError with code "lease_held". The lease is renewed in the background and handed over when you call release(), leave the with block, or close the client.

If the process before crashed without releasing, resume has to wait for its lease to expire. When you cannot wait, as on a live call, pass defer_lease=True: resume returns the context at once without the lease, because reading cannot conflict with anyone. Writes then raise RecallError with code "lease_not_held", so keep them in a buffer and call agent.acquire() until it returns True:

agent = client.resume("call-42", defer_lease=True, lease_ttl=6)
greet_with(agent.context.briefing)  # no waiting
while not agent.acquire():          # False while the crashed process's lease is live
    time.sleep(1)
agent.record_turn("assistant", "Sorry, we got cut off.")

Once it holds the lease, anything the crashed process wrote after the resume is picked up, and new turns continue after it.

The other methods write what the next resume reads: update_working_state (fields you leave out are kept), milestone, record_turn (a turn longer than the output threshold is stored whole and the turn keeps a reference), store_output and fetch_output for large tool results, set_pointer, remove_pointer and pointers, and recent_turns and working_state_history to read back.

Tool calls that touch the world

The records above say what the agent was doing, not whether its last tool call happened. Record each call that changes something before it runs, and close it when it returns:

action = agent.begin_action("send_email", "external", {"to": to, "subject": subject})
result = mailer.send(to, subject)
agent.end_action(action.seq, "done", result=result)

The class says what running the call again would do: "pure" (reads only), "idempotent" (repeating it with action.key does nothing more, so pass the key on) or "external" (it could happen twice). If the process dies between begin_action and end_action, the call is still open on the next resume. Interrupted pure calls are closed for you and external ones are marked unknown. Every call still open is in context.pending_actions and at the top of the briefing, as one to check before repeating.

agent.reconcile(resolver) settles them. The resolver gets each open action: for an idempotent one it runs the call again with action.key; for an external one it asks the outside system whether it happened. It returns ActionEnd("done") or ActionEnd("failed"), or None when it cannot tell. Each settled outcome is written to the agent's history, so the model sees it, and context is rebuilt. Recall never runs a tool itself.