feat(memory): add sync_recall option for current-turn relevance

[heathley]

Summary

Closes #5820.

Adds a memory.sync_recall: true config option that switches memory providers from background prefetch to synchronous recall at turn start.

Problem

The current model queues a background recall at turn end using that turn's query. The next turn consumes this result — which is based on the previous message. When topics change between turns the injected context is irrelevant or misleading.

Solution

When sync_recall: true, each turn performs a live recall with the actual current message before the first LLM call. Background prefetch is skipped entirely at turn end. Default is false — no behavior change for existing users.

Implementation

MemoryProvider.recall_sync() default: fires queue_prefetch() then immediately calls prefetch(), which joins the background thread. This works for all providers that join inside prefetch() (honcho, hindsight, mem0, openviking). RetainDB overrides recall_sync() directly because its prefetch() reads shared state without joining threads.

4 files changed, 54 lines added.

Config

memory:         
  provider: hindsight  # or honcho, mem0, etc.
  sync_recall: true

[teknium1]

Thanks for tackling the stale/empty first-turn recall problem. The premise remains present on current main: agent/turn_context.py:558-564 calls prefetch_all() before the tool loop, and Hindsight consumes a queued result in plugins/memory/hindsight/__init__.py:1466-1484.

Problems

• The generic recall_sync() is not synchronous for current Honcho. plugins/memory/honcho/session.py:666-679 starts base-context work in a daemon thread, while plugins/memory/honcho/session.py:688-695 only returns a ready cached result. queue_prefetch() followed immediately by prefetch() can therefore still inject stale or empty base context.
• Current main moved the pre-turn integration to agent/turn_context.py:558-564; the PR's agent/conversation_loop.py edit needs a current-tree port. The new manager path should also preserve prefetch_all()'s skill-scaffolding normalization at agent/memory_manager.py:501-503.
• The PR changes no tests. Please add behavioral coverage for current-query recall, first-turn behavior, post-turn prefetch suppression, and provider-specific paths.

Suggested changes

• Give Honcho an explicit synchronous recall implementation or a provider contract that can actually satisfy this guarantee.
• Port the feature through turn_context.py, retain normalization, and add the config default/documentation.

Automated hermes-sweeper review.

[teknium1]

This generic composition is not synchronous for current Honcho: its prefetch_context() starts a daemon thread and its prefetch() only consumes a ready cache (plugins/memory/honcho/session.py:666-695). Please provide an explicit synchronous Honcho path (or narrow this contract) before claiming current-query relevance.

[teknium1]

When porting onto current main, normalize this query exactly as prefetch_all() does at agent/memory_manager.py:501-503; otherwise /skill and bundle-expanded turns send provider recall the full injected skill body rather than the user instruction.

[kshitijk4poor]

Merged via #85494 — the sync_recall concept from this PR was incorporated as recall_sync in the Hindsight provider config schema. Thanks for the original idea!