Memory and payload lifecycle

ZeroClaw carries several kinds of “remembered” information during a turn. They do not all have the same owner, durability, privacy boundary, or review risk.

Use this page when a change touches memory, history, session persistence, tool results, files, media attachments, summaries, context trimming, or prompt assembly. The most important question is not “does the agent remember this?” but “which surface owns this data, and how long does it live?”

What owns what

Surface	Owner	Durability	What reviewers should check
Long-term memory	`zeroclaw-memory` behind `Arc<dyn Memory>`	Backend-specific: SQLite/Postgres/Lucid/Qdrant/shared stores, or per-agent Markdown files	Stores and recalls must stay agent-scoped. A tool result, log line, or session row is not long-term memory unless a memory write happened.
Relationship memory	`knowledge` tool and knowledge graph	Graph backend, when enabled	Capture is explicit. Enabling the graph does not automatically ingest conversations, files, or channel data.
Session history	`zeroclaw-infra` session backends, ACP store, and live RPC/session maps	Chat/ACP history can persist; live RPC handles are process-local	History preserves conversation continuity. It is not the canonical store for user preferences, config, or files.
Current prompt context	Agent loop prompt assembly	Ephemeral provider request	Recalled memory, hardware RAG, current input, system prompt, skills, and tool results may be sent to the provider. This does not make them durable.
History trimming	`agent::history` and `agent::history_trim`	Lossy change to the request/session history shape	Trimming must be visible, preserve tool-call/tool-result pairing, and avoid pretending old context is still available.
Tool result payloads	`ToolResult`, `ToolResultMessage`, and the tool dispatcher	Current turn and any persisted session history that records the turn	Bound size and provenance. Large outputs should be capped or summarized intentionally; image-path promotion must only happen for producing tools, not path-listing tools.
Files and workspaces	Per-agent workspace security policy	Files persist according to the filesystem, not memory	File contents are not memory just because a tool read them. Writes belong in the agent workspace unless policy explicitly allows more.
Media attachments	Channel/gateway media pipeline and `MediaAttachment`	Inbound payload by default; persistence depends on the receiving path	Raw bytes should stay bounded and path-validated. Store summaries or references deliberately rather than silently copying media into memory.
Logs and observer events	`zeroclaw-log`, `ObserverEvent`, runtime trace	Optional runtime trace and live observers	Logs are evidence and diagnostics, not source-of-truth memory. Scrub or bound user/tool payloads before logging.
Cost and usage records	Cost tracker and provider usage events	Cost ledger when enabled	Usage records describe model calls. They should not carry prompt bodies, tool outputs, or memory contents.

This table complements Runtime state and persistence. That page says where state lives; this page says how user-facing payloads move through memory, history, tools, files, media, and provider requests.

Long-term memory

An agent receives its memory handle from the memory factory. Shared backends and Markdown storage have different concrete layouts, but the review rule is the same: memory access must stay bound to the agent identity and configured peer allowlist described in Runtime internals.

There are two normal ways information becomes durable memory:

the agent calls a memory tool such as memory_store;
runtime code explicitly stores a memory entry, such as the configured conversation autosave path.

Do not treat prompt context, tool output, files, or logs as durable memory by default. A PR that makes one of those surfaces persistent must name the memory category, session scope, agent scope, retention behavior, and operator-visible control.

Prompt context and recall

At turn start, the runtime can recall relevant memories and inject a bounded [Memory context] block into the user-visible prompt context. Related entry points do not all apply identical filters. The channel/interactive loop filters generated autosave noise, stale <tool_result> blocks, and Conversation entries when the turn has no safe session scope or is not user-initiated. Generic memory loading filters autosave noise and relevance, but does not by itself enforce that channel-loop Conversation exclusion.

The provider request can therefore contain recalled memory without making the current turn a new memory. Review prompt-assembly changes by asking:

which memory backend and agent scope were queried;
whether the query is session-scoped when conversation entries are allowed;
whether autosave noise, stale tool-result blocks, and low-relevance entries remain filtered;
whether the user or operator can see when older context was removed.

Session history and trimming

Session history is the continuity record for a conversation. It can include chat messages, assistant tool calls, and tool results. It is not the same thing as long-term memory.

History management owns the trimming mechanics. This page only names the lifecycle boundary: trimming is a lossy change to provider-visible/session-visible context, not a memory delete, and it must be visible rather than silently pretending old context remains available.

Tool-call pairing matters more than byte savings. A history change must not leave a provider request with a dangling tool_use without the matching tool_result, or the reverse.

Tool results

Tools return a small structured result: success, output, and error. The dispatcher converts those results into provider messages for the next model call, while streaming clients can receive correlated ToolCall and ToolResult events during the turn.

Tool result payloads are easy to over-preserve. Reviewers should check:

maximum result size, including max_tool_result_chars;
whether truncation preserves structured envelopes and image markers;
whether search/listing tools avoid turning incidental image paths into media payloads;
whether receipts, logs, and observer events carry bounded, scrubbed evidence rather than raw sensitive output;
whether the result is only in the current turn/session history or is also intentionally written to memory.

If a PR says a tool result is “remembered”, require it to say whether that means provider-visible history, persisted session history, a memory backend row, a file artifact, a receipt, or a log event.

Files and media

File contents and media bytes are payloads, not memories. The filesystem owner is the per-agent workspace policy described in Filesystem components and Runtime internals. A file read can place content in a tool result or prompt; a file write can create durable filesystem state; neither automatically creates a memory row.

Inbound channel messages can carry MediaAttachment values with a file name, bytes, and optional MIME type. MediaKind is derived from MIME type or file extension. The low-level attachment loader reads caller-supplied paths verbatim, so callers that accept untrusted paths must validate or constrain those paths before loading.

For files and media, reviewers should look for:

workspace policy enforcement before reads and writes;
path validation when a path came from a user, HTTP request, channel payload, or tool argument;
bounded byte handling and clear failure behavior for missing or unreadable files;
explicit summaries or references when large/binary payloads enter prompts;
no silent copy from attachment or file content into long-term memory.

Logs and observability

Observer events and runtime logs help explain what happened. They should not become hidden payload stores. Memory recall events carry a scrubbed/truncated query summary and counts. Memory store events carry bounded category and backend identifiers.

Tool-call observability needs extra care because the sinks do not share one payload contract. Current typed tool-call observer events can carry full arguments and credential-scrubbed full result output, and OTel forwards those values into span attributes. Do not describe that path as “summaries” unless the code actually bounds or summarizes it. New telemetry should prefer bounded identifiers, counts, durations, success flags, and operator-useful summaries. Put raw content in logs or observer events only when the feature explicitly requires it and the privacy boundary is documented.

Reviewer checklist

For memory, payload, history, file, or media changes, answer these before reviewer sign-off:

What is the canonical owner of the data?
Is it current-turn only, session-persistent, filesystem-persistent, memory-persistent, or log-persistent?
Which agent, session, channel, or workspace scope limits access?
Can a caller widen memory recall beyond the configured allowlist?
Can autonomous jobs see chat-origin conversation memory?
What bounds tool output, file bytes, media bytes, and prompt size?
Does trimming or truncation make loss visible instead of silent?
Are provider-visible payloads separated from durable memory writes?
Are logs and observer events scrubbed and bounded?
If the PR changes a generated or derived payload, does it update the source owner rather than hand-editing generated output?

Source pointers

Canonical docs:

Key code entry points:

Memory trait and entry shape: crates/zeroclaw-api/src/memory_traits.rs
Memory factory and agent scoping: crates/zeroclaw-memory/src/lib.rs, crates/zeroclaw-memory/src/agent_scoped.rs, and crates/zeroclaw-memory/src/agent_scoped_markdown.rs
Memory tool registry and examples: crates/zeroclaw-tools/src/lib.rs (MEMORY_TOOL_NAMES), crates/zeroclaw-tools/src/memory_store.rs, and crates/zeroclaw-tools/src/memory_recall.rs
Prompt recall and injection: crates/zeroclaw-runtime/src/agent/memory_inject.rs (recall policy plus the [Memory context] renderer), injected engine-side in crates/zeroclaw-runtime/src/agent/turn/mod.rs; the per-turn memory handle is threaded through crates/zeroclaw-runtime/src/agent/loop_.rs
History trimming and tool-result payload shaping: crates/zeroclaw-runtime/src/agent/history.rs, crates/zeroclaw-runtime/src/agent/history_trim.rs, and crates/zeroclaw-runtime/src/agent/turn/results_collect.rs
Tool and provider message shapes: crates/zeroclaw-api/src/tool.rs and crates/zeroclaw-api/src/model_provider.rs
Channel attachments: crates/zeroclaw-api/src/channel.rs and crates/zeroclaw-api/src/media.rs

Keyboard shortcuts

ZeroClaw Docs