fix(claude-code): send MemoryItem.timestamp so the extractor can resolve relative dates - #2974

Of ~1,400 facts accumulated in a real bank, exactly 1 carried a usable date. The retain hook never sends MemoryItem.timestamp, so the extractor has no anchor: "yesterday" and "last week" are stored verbatim, occurred_start stays null, and the memory never reaches the Timeline. The hook already computes the value — it just never reaches the request body.

Problem

The Claude Code retain hook never sends MemoryItem.timestamp, so the extractor has no reference time to resolve relative expressions in a transcript.

The result is facts that are stored but not usable later. A session saying "yesterday the sensors phantom-cascaded" is extracted as:

occurred_start=None | ... | When: Today (relative to conversation)
occurred_start=None | ... | When: Last week

"Last week" relative to what? The anchor is exactly the information that was dropped. Once that fact is recalled weeks later it conveys nothing, and occurred_start stays null so the memory never reaches the Timeline.

This is listed as an anti-pattern in Hindsight's own best-practices guide:

Missing timestamp on retain — Disables temporal retrieval strategies — Always set from actual content timestamps

and in the timestamp field docs:

Set whenever you have temporal context. Enables temporal retrieval strategies.

Measurement

Same transcript, same retain_mission, run through dry-run-extract on a self-hosted server with the anchor as the only variable:

On a bank of ~1,400 facts accumulated without the anchor, exactly 1 carried an occurred_start.

Fix

The hook already computes the instant it needs. run_retain builds template_vars["timestamp"] (time.gmtime(), ISO 8601) and records it as retained_at metadata — it just never reached the request body. This passes that same value through as the MemoryItem reference time.

• No new configuration.
• No behaviour change for callers that omit it: HindsightClient.retain(timestamp=None) sends no timestamp key at all, so the server keeps its existing ingestion-time fallback rather than receiving an explicit null.

Tests

Two added to TestRetainHook (203 pass, up from 201):

• test_retain_sends_timestamp_for_relative_date_resolution — asserts the posted item carries an ISO-8601 timestamp and that it matches the retained_at metadata, so both describe the same instant.
• test_client_omits_timestamp_field_when_not_supplied — pins the omission path.

Mutation-checked: deleting the item["timestamp"] = timestamp assignment fails the first test; restoring it passes.

Scope note

This only supplies the anchor. Whether the extractor then populates occurred_start is model-dependent — a local qwen3.6 resolved relative dates into the fact text but still left occurred_start null, while a hosted model populated it 3/5 of the time on the same input. Sending the timestamp is a precondition for either, and improves the stored text regardless of which model is behind the extractor.