feat: Honcho memory integration (opt-in) by erosika · Pull Request #38 · NousResearch/hermes-agent

erosika · 2026-02-26T01:33:23Z

Cross-session user modeling via Honcho. Hermes builds a persistent representation of who it's talking to -- preferences, patterns, context -- and carries that understanding across conversations. A model of the user.

Three integration points: prefetch user context into the system prompt each turn, sync exchanges for ongoing modeling, and a dialectic reasoning tool (query_user_context) for the agent to actively query its understanding of the user mid-conversation.

honcho_integration/ reads ~/.honcho/config.json as single source of truth. All imports lazy, all calls non-fatal. Optional dependency, zero behavior change when disabled.

Composes with USER.md

Both systems run in tandem. USER.md stays in the system prompt as the agent's local, curated snapshot of the user. Honcho prefetch is additive -- synthesized cross-session context layered on top.

When the agent writes to memory(target="user", action="add"), observations go to both USER.md (local file) and Honcho (cross-session reasoning model). MEMORY.md (agent's own notes) is untouched in all cases.

When Honcho is disabled, USER.md works exactly as before. Nothing changes.

Prefetch size is configurable via contextTokens in the global config -- caps how much user context Honcho surfaces per turn. No value = uncapped.

erosika · 2026-02-26T01:34:57Z

teknium1 · 2026-02-26T07:41:04Z

Can you link me to more info on Honcho I havent heard of it to know what it adds vs the USER.md user profile that the agent already builds on their user

erosika · 2026-02-26T16:38:01Z

@teknium1

So USER.md is creating conclusions similar to how Honcho does, yes. To be clear, my PR currently runs alongside it -- I'll make one more commit to integrate them cleanly (still opt-in) if you'reinterested.
But with Honcho it's more than selective conclusions / RAG over said conclusions. You have configurable mapping of sessions, peers, and workspaces under the hood. I have ~100,000 conclusions about me captured via Claude Code over the last 3 months for example-- with Honcho, I can make that context immediately available to Hermes, compose chat turns for different levels of reasoning (.chat()) and summarization (.context()), and ingest that into the agent's immediate or long term awareness.

Important to note-- the reasoning Honcho does is powered by Neuromancer which was specifically fine-tuned for this task. (link to evals here)

There's also dreaming -- the reasoning model ambiently makes conclusions and fills in the gaps between sessions. It's a better way to form identity, all accessible in the cloud (remap anything on the fly) or self-hosted with your own model infra.

I'm using Honcho as a bridge between all of my agentic interactions and that's the most valuable part for me. The peer-peer architecture also lets you control which peers can make conclusions about other peers -- super helpful for agentic orchestration going forward.

Happy to discuss more -- erosika on Discord / @3rosika on Twitter.

Announcing Honcho 3 | honcho.dev

Opt-in persistent cross-session user modeling via Honcho. Reads ~/.honcho/config.json as single source of truth (shared with Claude Code, Cursor, and other Honcho-enabled tools). Zero impact when disabled or unconfigured. - honcho_integration/ package (client, session manager, peer resolution) - Host-based config resolution matching claude-honcho/cursor-honcho pattern - Prefetch user context into system prompt per conversation turn - Sync user/assistant messages to Honcho after each exchange - query_user_context tool for mid-conversation dialectic reasoning - Gated activation: requires ~/.honcho/config.json with enabled=true

When Honcho is active: - System prompt uses Honcho prefetch instead of USER.md - memory tool target=user add routes to Honcho - MEMORY.md untouched in all cases When disabled, everything works as before. Also wires up contextTokens config to cap prefetch size.

USER.md stays in system prompt when Honcho is active -- prefetch is additive context, not a replacement. Memory tool user observations write to both USER.md (local) and Honcho (cross-session) simultaneously.

teknium1 · 2026-02-28T07:33:32Z

Okay, after reviewing I like it a lot. I'll add more documentation on it after merge - Thanks!

…e integration, setup CLI Authored by erosika. Builds on #38 and #243. Adds async write support, configurable memory modes, context prefetch pipeline, 4 new Honcho tools (honcho_context, honcho_profile, honcho_search, honcho_conclude), full 'hermes honcho' CLI, session strategies, AI peer identity, recallMode A/B, gateway lifecycle management, and comprehensive docs. Cherry-picks fixes from PRs #831/#832 (adavyas). Co-authored-by: erosika <erosika@users.noreply.github.com> Co-authored-by: adavyas <adavyas@users.noreply.github.com>

… files

a-canary · 2026-04-10T19:05:14Z

Investigated and resolved. The pi-bridge service is running correctly with the stdio JSON Director handler at (v0.3.0). The .env correctly specifies . An orphaned pre-ESM-migration Director tmux session was also cleaned up. Closing as resolved - the system is operational.

a-canary · 2026-04-10T19:05:21Z

Investigated and resolved. The pi-bridge service is running correctly with the stdio JSON Director handler (v0.3.0). The .env correctly specifies the path to director-bridge.js. An orphaned pre-ESM-migration Director tmux session was also cleaned up. Closing as resolved - the system is operational.

…gration feat: Honcho memory integration (opt-in)

…session title integration, setup CLI Authored by erosika. Builds on NousResearch#38 and NousResearch#243. Adds async write support, configurable memory modes, context prefetch pipeline, 4 new Honcho tools (honcho_context, honcho_profile, honcho_search, honcho_conclude), full 'hermes honcho' CLI, session strategies, AI peer identity, recallMode A/B, gateway lifecycle management, and comprehensive docs. Cherry-picks fixes from PRs NousResearch#831/NousResearch#832 (adavyas). Co-authored-by: erosika <erosika@users.noreply.github.com> Co-authored-by: adavyas <adavyas@users.noreply.github.com>

…gration feat: Honcho memory integration (opt-in)

…working) Deploy NousResearch#38's captured probe error named the cause exactly: from inside the gateway container, "ssh: connect to host 127.0.0.1 port 22: Connection refused" — the prod compose runs the containers on a BRIDGE network (as the health-gate comments already document), so the container's loopback is its own, not the host's. The host-side denial probe connecting fine confirms sshd is up on the host. Fix: - docker-compose.prod.yml: add extra_hosts host.docker.internal:host-gateway on the gateway service and set HERMES_FLEET_HOST=host.docker.internal, so the in-container fleet client targets the host's gateway address. Correct under both bridge and host networking. - deploy bootstrap: pin the host's public host keys under BOTH names (host.docker.internal for bridge-networked prod, 127.0.0.1 for host-networking setups) in the volume known_hosts. If the next run still can't connect, the captured probe error will say whether it's now a firewall filter on the docker subnet rather than a wrong address. https://claude.ai/code/session_018vp1LDdpoaRLCNZQRsnAp8

mrjohnyhats

Hermes Agent Review

Reviewed from first principles against the PR spec: opt-in Honcho, ~/.honcho/config.json as the source of truth, non-fatal optional dependency, additive USER.md behavior, prefetch/sync/query integration. I found two correctness/privacy blockers and one config-precedence bug. Since this PR is already merged, these should be follow-up fixes rather than merge blockers here.

Automated check run locally: python -m compileall honcho_integration tools/honcho_tools.py run_agent.py model_tools.py toolsets.py hermes_cli/config.py gateway/run.py passed.

mrjohnyhats · 2026-06-10T08:09:08Z

+_session_key: str | None = None  # Current session key (e.g., "telegram:123456")
+
+
+def set_session_context(session_manager, session_key: str) -> None:


⚠️ Privacy/correctness issue: this stores the active Honcho manager/session key in module-level globals. In gateway/API usage Hermes can have more than one AIAgent alive at a time (different chats, platforms, or overlapping tool loops). The last agent to call set_session_context() wins, so query_user_context can answer using another user's/session's Honcho context, and the availability check_fn can also be stale. This violates the user-model isolation that the feature needs. Please make the tool handler receive per-call agent/session context (or register a per-agent closure/context object) rather than shared module state.

mrjohnyhats · 2026-06-10T08:09:08Z

+        path = config_path or GLOBAL_CONFIG_PATH
+        if not path.exists():
+            logger.debug("No global Honcho config at %s, falling back to env", path)
+            return cls.from_env()


⚠️ Spec mismatch / opt-in regression: the PR description says activation is gated by ~/.honcho/config.json with enabled=true and that the global Honcho config is the single source of truth. But when the config file is absent or invalid this falls back to from_env(), which sets enabled=True. That means any process with HONCHO_API_KEY in the environment will start syncing conversations to Honcho without an explicit enabled=true config. Please keep env as an API-key fallback only after an explicit config opt-in, or have from_env() default to disabled unless a dedicated opt-in env/config flag is present.

mrjohnyhats · 2026-06-10T08:09:08Z

+            linked_hosts=linked_hosts,
+            enabled=raw.get("enabled", False),
+            save_messages=raw.get("saveMessages", True),
+            context_tokens=raw.get("contextTokens") or host_block.get("contextTokens"),


💡 Config precedence bug: the module docstring/spec says host-specific settings win over flat/global settings, but contextTokens is resolved in the opposite order (raw before host_block). If a user sets a global budget and a Hermes-specific override, the Hermes override is ignored. This should mirror workspace/aiPeer precedence: host_block.get("contextTokens") or raw.get("contextTokens").

…ault auf die kanonische Deploy-Linie portiert Portiert die einzige (uncommittete) Kopie des FAIL_CLOSED_PLATFORMS-Kerns von der veralteten Linie jarvis/voice-background-control (789 Commits hinter Deploy) auf jarvis/engine-deploy (= deployte d4f501d). Schliesst den fail-OPEN-Default fuer die untrusted Voice-Bahn (api_server): unklarer Config-Zustand -> leeres Toolset (no_mcp) statt vollem Toolset; MCP-Suppression spannt jetzt beide Entscheidungsstellen. Verifiziert: pytest 122 (Ziel-Dateien) + 559 (Regressions-Sweep toolset/mcp/config) gegen den gepatchten Code im Engine-Image, 0 Fehler. Reine Defense-in-Depth, kein Happy-Path-Verhalten geaendert (nur fail-open->fail-closed auf dem abnormalen Pfad). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…t-Enable + kanonischen MCP-Alias Schliesst zwei vom adversarialen 3-Lens-Review gefundene Luecken im fail-closed Boden (galten teils schon vor dem Port; der Port-Kommentar behauptete faelschlich 'general form'): V1 (high): der Plugin-Default-Enable-Block konsultierte FAIL_CLOSED_PLATFORMS NICHT -> ein gebuendeltes Plugin-Toolset nicht in _DEFAULT_OFF (z.B. google_meet) wurde auf dem [no_mcp]-Produktions-Pfad mit 0 Opt-in default-aktiviert. Fix: FAIL_CLOSED-Plattformen aktivieren Plugin-Toolsets nie per Default (nur explizit; Drift faengt Lage 2). V2 (broken): suppress_mcp zog nur Bare-Server-Namen ab, nicht den kanonischen 'mcp-<server>'-Toolset-Alias -> explizite ['mcp-foo']-Config leakte die volle MCP-Tool- Surface auf die Voice-Bahn. Fix: fuer FAIL_CLOSED_PLATFORMS auch 'mcp-*'-Aliase abziehen. SCOPED auf FAIL_CLOSED (nicht jedes suppress_mcp), damit die no_mcp-Sentinel-Semantik fuer cli/discord upstream-konform bleibt -- keine stille Divergenz im 1274-Commits-hinter-Fork. Scope bewusst: Auto-Enable-Suppression vervollstaendigt, KEIN Capability-Clamping expliziter Configs (bleibt Lage 2 Hash-Refuse-to-boot). Tests: +5 (2 lasttragend auf ROHER _get_platform_tools-Ausgabe + 3 Kontrollen/Scope-Guards); Revert-Check beweist die 2 Fix-Tests fallen bei neutralisiertem Guard. Voller Lauf: 564 passed (Ziel + Regressions-Sweep toolset/mcp/config), 0 Fehler. 2 adversariale Review-Runden: V1/V2 = closed (sound). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (hart gepinnt, inert bis Trigger) Konsumiert RESEARCH_ENQUEUE_DIR/*.json (schema research.enqueue.v1, von der Bridge nach /verify ok atomar geschrieben) und feuert EINEN bewachten Recherche-Job. HART GEPINNT, nie aus Datei/Input: enabled_toolsets=["web"], model=gpt-5.4, provider=openai-codex, repeat=1 (Eskalations-Schutz web->action). Schreibt das von der Bridge gepollte Format (research_index + _progress.json). Rekursions-Sperre (web-only + In-Prompt-Guardrail, kein Brain-Tool -> no_mcp bewahrt). Idempotenz: claim-before-fire (atomares Rename). DLP: topic/goal nie geloggt. Review-Fixes: unabhaengige per-Owner-Kostendecke (rate_ledger.jsonl, kein Art.9) + Retention- Reaper (Art.9-latente Terminal-Dateien nach TTL scrubben; rate-Signal bleibt). 60 Tests gruen. INERT bis: (a) periodischer run_once()-Trigger verdrahtet + (b) profile-Param-Wiring auf der engine-deploy-cronjob()-Linie (NousResearch#38-Folgearbeit) + Deploy. Datei-disjunkt (nur neue Worker/Test). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…gration feat: Honcho memory integration (opt-in)

…session title integration, setup CLI Authored by erosika. Builds on NousResearch#38 and NousResearch#243. Adds async write support, configurable memory modes, context prefetch pipeline, 4 new Honcho tools (honcho_context, honcho_profile, honcho_search, honcho_conclude), full 'hermes honcho' CLI, session strategies, AI peer identity, recallMode A/B, gateway lifecycle management, and comprehensive docs. Cherry-picks fixes from PRs NousResearch#831/NousResearch#832 (adavyas). Co-authored-by: erosika <erosika@users.noreply.github.com> Co-authored-by: adavyas <adavyas@users.noreply.github.com>

…click through (NousResearch#38) `artifact.action.invoke` for an `artifact.session.spawn` intent returned the short 8-char runtime session id. But the client resolves a spawned session against `session.list` rows, which carry only the long database id (`YYYYMMDD_HHMMSS_xxxxxx`) — and because the session is created server-side, the client never ran `session.create` to learn the runtime→db mapping. So the click-through silently failed to resolve. `_spawn_session` now seeds the task against the runtime id (the correct handle for in-memory `prompt.submit` dispatch) but returns `stored_session_id`, the stable key the client can match. Falls back to the runtime id when no database id is present. Tests: assert the DB id flows back to the client, the runtime id is used for the seed submit, and the fallback path returns the runtime id when no db id. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…rshell-output [codex] 修复 Windows PowerShell 终端对象输出丢失

…2-curator (NousResearch#38/NousResearch#40)

erosika added 3 commits February 26, 2026 18:07

refactor: run Honcho and USER.md in tandem

70d1abf

USER.md stays in system prompt when Honcho is active -- prefetch is additive context, not a replacement. Memory tool user observations write to both USER.md (local) and Honcho (cross-session) simultaneously.

erosika force-pushed the feat/honcho-integration branch from 6430f5b to 70d1abf Compare February 26, 2026 23:07

Merge branch 'main' into feat/honcho-integration

4a9086b

teknium1 closed this Feb 28, 2026

teknium1 reopened this Feb 28, 2026

teknium1 merged commit 912ef50 into NousResearch:main Feb 28, 2026

This was referenced Mar 9, 2026

feat(honcho): async writes, memory modes, session title integration, setup CLI #736

Merged

[Bug]: hermesagent consistently forgets he/she/it has shell access #747

Closed

h4x3rotab pushed a commit to Clawdi-AI/hermes-agent that referenced this pull request Apr 10, 2026

Merge PR NousResearch#38: fix memory search to include root-level .md…

0b28bdb

… files

kskotetsu mentioned this pull request Apr 18, 2026

MMC3 scanline IRQ実装バグ (#12テストがFAIL) #12147

Closed

angelburgosrosado pushed a commit to angelburgosrosado/hermes-agent that referenced this pull request Apr 27, 2026

Merge pull request NousResearch#38 from plastic-labs/feat/honcho-inte…

d7a0338

…gration feat: Honcho memory integration (opt-in)

gizdusum pushed a commit to gizdusum/hermes-agent that referenced this pull request May 17, 2026

Merge pull request NousResearch#38 from plastic-labs/feat/honcho-inte…

8700c5e

…gration feat: Honcho memory integration (opt-in)

onedayaname mentioned this pull request May 26, 2026

[Bug]: terminal tool hangs 50-65s when reusing existing session in CLI/TUI #32460

Open

mrjohnyhats reviewed Jun 10, 2026

View reviewed changes

ceo-aivocado mentioned this pull request Jun 16, 2026

[codex] Dispatch Telegram link mentions ceo-aivocado/hermes-agent#6

Merged

waefrebeorn pushed a commit to waefrebeorn/slermes that referenced this pull request Jul 2, 2026

Merge pull request NousResearch#38 from plastic-labs/feat/honcho-inte…

ba47913

…gration feat: Honcho memory integration (opt-in)

nicezic pushed a commit to nicezic/hermes-agent that referenced this pull request Aug 26, 2026

Merge pull request NousResearch#38 from Eynzof/codex/fix-windows-powe…

b8eed57

…rshell-output [codex] 修复 Windows PowerShell 终端对象输出丢失

Soju06 added a commit to Soju06/hermes-agent that referenced this pull request Aug 27, 2026

docs(patches): register 5-tuple arity fixes under memory-phase1/phase…

5d3bdae

…2-curator (NousResearch#38/NousResearch#40)

Navigation Menu

Uh oh!

feat: Honcho memory integration (opt-in) - #38

feat: Honcho memory integration (opt-in)#38
teknium1 merged 4 commits into
NousResearch:mainfrom
plastic-labs:feat/honcho-integration

erosika commented Feb 26, 2026 •

edited

Loading

Uh oh!

erosika commented Feb 26, 2026

Uh oh!

teknium1 commented Feb 26, 2026

Uh oh!

erosika commented Feb 26, 2026 •

edited

Loading

Uh oh!

teknium1 commented Feb 28, 2026

Uh oh!

a-canary commented Apr 10, 2026

Uh oh!

a-canary commented Apr 10, 2026

Uh oh!

mrjohnyhats left a comment

Uh oh!

mrjohnyhats Jun 10, 2026

Uh oh!

mrjohnyhats Jun 10, 2026

Uh oh!

mrjohnyhats Jun 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

		_session_key: str \| None = None # Current session key (e.g., "telegram:123456")


		def set_session_context(session_manager, session_key: str) -> None:

Uh oh!

Conversation

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

Uh oh!

Composes with USER.md

Uh oh!

erosika commented Feb 26, 2026

Uh oh!

teknium1 commented Feb 26, 2026

Uh oh!

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

Uh oh!

Uh oh!

teknium1 commented Feb 28, 2026

Uh oh!

a-canary commented Apr 10, 2026

Uh oh!

a-canary commented Apr 10, 2026

Uh oh!

mrjohnyhats left a comment

Choose a reason for hiding this comment

Hermes Agent Review

Uh oh!

mrjohnyhats Jun 10, 2026

Choose a reason for hiding this comment

Uh oh!

mrjohnyhats Jun 10, 2026

Choose a reason for hiding this comment

Uh oh!

mrjohnyhats Jun 10, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

erosika commented Feb 26, 2026 •

edited

Loading

erosika commented Feb 26, 2026 •

edited

Loading