feat(memory): improve OpenViking runtime reliability by ehz0ah · Pull Request #32444 · NousResearch/hermes-agent

ehz0ah · 2026-05-26T06:32:16Z

Details

This improves how the Hermes OpenViking memory provider retrieves, writes, and commits memory during normal agent runs.

Behavior changes

Recall/prefetch now fetches current-query context when the cached background result is missing or stale.
Retrieved OpenViking candidates are deduplicated, reranked for query overlap, enriched with bounded detail reads, and injected as a capped evidence block instead of unbounded raw results.
Session writes now go through a single async worker, use the canonical message-parts payload, retry once with a fresh client after write failure, and are flushed before session commit. Commit is skipped when queued writes do not drain or when the OpenViking session has no pending tokens.
Hermes memory add operations and explicit viking_remember tool calls write durable memories via /api/v1/content/write; normal turn transcripts remain session messages for commit-time extraction.
OpenViking tools keep outputs bounded, and viking_read can read a small batch of URIs for efficient follow-up.

Test Plan

scripts/run_tests.sh tests/plugins/memory/test_openviking_provider.py tests/agent/test_memory_provider.py tests/agent/test_memory_async_sync.py tests/agent/test_memory_session_switch.py
.venv/bin/python -m ruff check plugins/memory/openviking/init.py tests/plugins/memory/test_openviking_provider.py
git diff --check

kshitijk4poor · 2026-06-24T13:28:43Z

Merged via #51898. The useful parts of your mixed runtime draft — synchronous current-query prefetch and the recall/prefetch direction — were folded into the consolidated PR. Thanks for the work.

@ehz0ah

Salvage of PR #48927 by @ehz0ah, which consolidates OpenViking recall work from #41706 (@huangxun375-stack), #33260, #49975, and #32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR #48927 had removed them; #41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

@ehz0ah

Salvage of PR NousResearch#48927 by @ehz0ah, which consolidates OpenViking recall work from NousResearch#41706 (@huangxun375-stack), NousResearch#33260, NousResearch#49975, and NousResearch#32444. Replaces stale background post-turn prefetch warming with synchronous current-query recall. The old queue_prefetch warmed the PREVIOUS user message while turn-start recall consumed the CURRENT one, so injected context was always about the wrong topic. Changes: - prefetch() now does session-aware /api/v1/search/search with the current query, falls back to /api/v1/search/find on failure - Contract-safe payloads: limit, score_threshold, context_type, session_id — no top_k, no search-body mode, no target_uri - L2 content reads for items with level=2 or empty abstracts, capped at full_read_limit (default 2) - Local ranking (score + query-token overlap + leaf boost), dedup, score threshold, and injected-char budget - queue_prefetch() is now a no-op (background warming removed) - Additive batched viking_read: uris param accepts up to 3 URIs - Per-request timeout support on _VikingClient.get/post/delete - Removes stale _prefetch_result/_prefetch_thread/_prefetch_generation state and _invalidate_prefetch_state() - Strengthened system_prompt_block guidance Salvage follow-up fixes: - Expose all 8 recall config knobs in get_config_schema() (PR NousResearch#48927 had removed them; NousResearch#41706 correctly exposed them). Env vars remain as internal mechanism but are now visible in setup wizard. - Lower default timeout 8s→4s, request_timeout 6s→3s, full_read_limit 3→2 to reduce per-turn blocking latency. Co-authored-by: Hao Zhe <haozhe4547@gmail.com> Co-authored-by: Eurekaxun <eurekaxun@163.com>

This was referenced May 26, 2026

feat(memory): improve OpenViking reliability and setup UX #24256

Closed

feat(memory): improve OpenViking setup UX #32445

Closed

alt-glitch added type/feature New feature or request P3 Low — cosmetic, nice to have comp/plugins Plugin system and bundled plugins tool/memory Memory tool and memory providers labels May 26, 2026

alt-glitch mentioned this pull request May 27, 2026

fix(memory): synchronous OpenViking prefetch with current turn's query #33260

Closed

ehz0ah force-pushed the feat/ov-runtime-optimisation branch from 96f2ee1 to a42a7d0 Compare June 15, 2026 07:38

ehz0ah added 4 commits June 16, 2026 21:52

feat(memory): improve OpenViking recall and write reliability

52c7642

refactor(memory): keep OpenViking prefetch query generic

c6370aa

fix(memory): preserve OpenViking remember session writes

9560c3e

fix(memory): separate OpenViking memory and session writes

1514c24

ehz0ah force-pushed the feat/ov-runtime-optimisation branch from a42a7d0 to 1514c24 Compare June 16, 2026 13:53

test(memory): cover OpenViking write edge cases

60345ad

ehz0ah mentioned this pull request Jun 17, 2026

fix(openviking): harden session switching and writes #47662

Closed

14 tasks

alt-glitch mentioned this pull request Jun 21, 2026

fix(openviking): synchronous prefetch with current turn's query #49889

Closed

ehz0ah marked this pull request as draft June 22, 2026 02:26

ehz0ah mentioned this pull request Jun 23, 2026

feat(openviking): add full recall prefetch policy #48927

Closed

19 tasks

ehz0ah closed this Jun 23, 2026

kshitijk4poor mentioned this pull request Jun 24, 2026

feat(openviking): add full recall prefetch policy (salvage #48927) #51898

Merged

Navigation Menu

feat(memory): improve OpenViking runtime reliability - #32444

feat(memory): improve OpenViking runtime reliability#32444
ehz0ah wants to merge 5 commits into
NousResearch:mainfrom
ehz0ah:feat/ov-runtime-optimisation

ehz0ah commented May 26, 2026 •

edited

Loading

Uh oh!

kshitijk4poor commented Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

ehz0ah commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Details

Behavior changes

Test Plan

Uh oh!

kshitijk4poor commented Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ehz0ah commented May 26, 2026 •

edited

Loading