Hermes Agent
Hermes proposes optional raw Mem0 sync after an exact-recall trade-off test
A Hermes proposal makes Mem0's automatic extraction configurable without changing the default. Its author reports faster ingestion and stronger literal-token recall with extraction disabled, but explicitly warns that the benchmark favors exact strings over summarized facts. The switch remains unmerged.
A September 22 Hermes proposal would let Mem0 users choose whether automatic per-turn synchronization runs an extraction model. The important restraint is that the default stays on. PR #119629 initially arose from different behavior between the explicit memory-add tool and background synchronization: explicit adds used infer=False, while per-turn sync used infer=True. The author found a test deliberately pinning that default and proposed a setting rather than silently overturning it.
The new sync_infer key is read from the provider's mem0.json, with absent values preserving current behavior. The inspected diff passes that value into the existing write path and adds an opt-out test. It does not replace the memory backend, change the explicit add tool, or prove that extraction is generally harmful.
The accompanying local experiment used 607 messages drawn from 31 sessions and 39 questions keyed to distinctive tokens. With extraction disabled, reported recall at five results was 0.9487 and ingestion took 7.7 seconds; with extraction enabled, recall was 0.4103 and ingestion took 1,187.1 seconds. The author explicitly acknowledges that this scorer disadvantages summaries, which can accurately express a fact without retaining its exact identifying token.
The facts
- The proposed sync_infer key defaults to true and belongs to the Mem0 provider's own configuration. - Automatic synchronization currently requests extraction, while the explicit memory-add path does not. - The author's comparison used Mem0 2.1.0, on-disk Qdrant, a local 768-dimensional embedder and an extraction model. - The dataset contained 607 session messages and 39 exact-token questions, not a general reasoning benchmark. - The author reports 32 passing Mem0 tests, including the unchanged default assertion and a new opt-out test. - The PR was open on inspection; the source does not establish a released configuration capability.
Why it matters
Memory systems optimize different things: readable distilled facts, exact identifiers, ingestion cost and retrieval usefulness are not interchangeable objectives. A default extraction pass can be reasonable for prose preferences but less suitable for a workload that must recover literal strings. Exposing that decision is more defensible than presenting one small experiment as a universal ranking of memory approaches.
Current
Checked September 23, 2026. The inspected release baseline is v2026.9.21. The primary source is open proposal. Tagged release availability and development-source state are separate: a release baseline does not establish that a proposed setting or newly merged feature is in an installed package.
Evidence
Primary: NousResearch/hermes-agent PR #119629 (https://github.com/NousResearch/hermes-agent/pull/119629). Context and implementation: NousResearch/hermes-agent issue #119627 (https://github.com/NousResearch/hermes-agent/issues/119627); NousResearch/hermes-agent — __init__.py (source) (https://github.com/NousResearch/hermes-agent/blob/7ed6327473520d0111ccae0679fcb31927781e1f/plugins/memory/mem0/__init__.py). The linked source text was inspected; reported experiments and test results remain attributed to their authors.
Operator take
For evaluation, use a redacted workload containing both exact identifiers and paraphrased questions. Compare storage size, ingestion latency, model calls, retrieval quality and retention policy; do not pick a setting solely from recall at five. Raw synchronization can preserve details that summarization removes, which also makes it a data-minimization choice. Keep the benchmark's model, corpus and scorer attached to any performance claim so later configuration changes do not inherit unrelated numbers.
Caveat
The benchmark and test counts are contributor reports, not independently reproduced results. Literal-token scoring favors raw retention, and the author says extracted facts remained accurate and readable. The setting is proposed, defaults to the existing extraction behavior and carries no production performance guarantee.
The benchmark and test counts are contributor reports, not independently reproduced results. Literal-token scoring favors raw retention, and the author says extracted facts remained accurate and readable. The setting is proposed, defaults to the existing extraction behavior and carries no production performance guarantee.