Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
2.7 KB
Raw Blame

Time-aware recall

gbrain think answers in a date frame. It knows today's date in your brain's timezone, it sees the content date of each page it reads, and it resolves relative time words against the right one: "last month" in your question against today, "yesterday" inside a meeting note against that note's date.

Say to your agent

  • "What did I decide about pricing last month?" — your agent runs gbrain think "What did I decide about pricing last month?"; the answer is grounded in today's date and each page's date.
  • "Answer as if today were March 1, 2024." — gbrain think "…" --reference-date 2024-03-01.

What the reader sees

  • Current date. The user message carries Current date: YYYY-MM-DD (<zone>) just before the question. The zone is brain.timezone (gbrain config set brain.timezone America/Los_Angeles); unset means UTC. --reference-date / MCP reference_date replaces today with a past or current YYYY-MM-DD; a malformed or future date is refused with invalid_params before any model call.
  • Page dates. Each <page> block carries date="YYYY-MM-DD" when the page has a content date: frontmatter event_date, date or published, or a dated filename. A page whose date fell back to when the file or row was created carries no date, because that time says nothing about the content. A day-only frontmatter date renders as written; a timestamp renders in the brain's timezone.
  • Search results already carry each page's effective_date and effective_date_source, so an agent reading search or query output directly can apply the same rule.

Measured effect

On held-out LoCoMo conversations, think answered 88.2% of questions correctly with the date frame against 74.2% without it (+14.0 points, 95% CI [+11.8, +16.2]); questions about time went from 27 to 199 of 221 correct. p95 latency is unchanged. On the LongMemEval-S development sample, think answered 90.0% of questions correctly with the date frame against 80.7% without it (+9.3 points, 95% CI [+4.0, +15.3]); on the LoCoMo development conversations, 89.1% against 76.7%. Retrieval is unchanged. Results: docs/eval/TIME_AWARE_RETRIEVAL_RESULTS.md.

Measuring it

The LongMemEval harness carries eval-only arms for two retrieval mechanisms that were measured and not shipped: fact keys merged into chunk embeddings (--fact-keys) and a soft time scope driven by the question's explicit time words (--time-scope). --mode tokenmax builds production per-chunk synopses, and scripts/locomo-to-longmemeval.ts converts LoCoMo conversations to the same format. See docs/eval-bench.md and the evaluation key files.