docs(harness-hermes): 0.18.x memory/skill-loop verdict + Engram boundary (OPS-480)

docs/eval/2026-07-07-model-harness-landscape-survey.md

@@ -16,6 +16,14 @@ Research project now reported as the fastest-growing agent framework in open
 source (60k+ stars in two months) on the strength of persistent multi-tier
 memory + a skill-writing learning loop — a pin-audit is Top-5 item 3.
 
+> **CORRECTION 2026-07-09 (OPS-480, source-verified):** the "persistent
+> multi-tier memory" claim above does not hold up against
+> `hermes_cli/config.py` on `NousResearch/hermes-agent@main` — the built-in
+> memory provider is two flat Markdown files (`MEMORY.md`, `USER.md`) with no
+> tier granularity, not three tiers. The skill-writing loop is confirmed
+> present and on by default at 0.18.0. See
+> `packages/harness-hermes/README.md` Verdict section for the full audit.
+
 ## 1. Models for the cluster
 
 ### Pragmatic — fits gateway upstreams or 32GB local tier now
@@ -88,6 +96,15 @@ memory + a skill-writing learning loop — a pin-audit is Top-5 item 3.
   60k+ stars in two months. **We pin 0.18.x — audit whether the memory/skill
   loop postdates the pin.** [site](https://hermes-agent.org/) ·
   [writeup](https://www.tencentcloud.com/techpedia/143930)
+  > **CORRECTION 2026-07-09 (OPS-480, source-verified):** memory is two flat
+  > Markdown files (`MEMORY.md` agent notes, `USER.md` user profile) injected
+  > into the system prompt, plus an empty-by-default external-provider slot
+  > (7 options, none of them Engram) — NOT three-tier semantic/working/
+  > episodic; that framing doesn't correspond to any config surface at
+  > 0.18.0. Skill-writing loop confirmed present and on by default, with
+  > unrestricted writes (`skills.write_approval: false` upstream default).
+  > Full audit + Engram-boundary lane decision in
+  > `packages/harness-hermes/README.md` Verdict section.
 - **OpenClaw** (ex-Moltbot/Clawdbot) — 280k+ stars (reportedly passed React;
   9k→60k in 72h at peak). Local-first gateway, 20+ messaging channels,
   cron/nodes/canvas first-class, 13.7k-skill marketplace, Markdown memory.

packages/harness-hermes/README.md

@@ -41,3 +41,96 @@ unsigned-paas runs this same upstream (pinned 0.16.0) as a kagent BYO agent
 with an A2A shim (`unsigned-paas/applications/hermes-agent*`). This package is
 the *local/workstation* harness; bumping the cluster pin to 0.18.0 is tracked
 in Linear, not done here.
+
+## Verdict — 0.18.x memory/skill-loop review (OPS-480, 2026-07-08)
+
+Review-only: **no version bump.** The 2026-07-07 landscape survey
+(`docs/eval/2026-07-07-model-harness-landscape-survey.md`) flagged hermes-agent
+as fastest-growing on the strength of "three-tier persistent memory
+(semantic/working/episodic) + a closed skill-writing learning loop" and asked
+whether that postdates our 0.18.0 pin. Findings below come from reading
+`hermes_cli/config.py` on `NousResearch/hermes-agent@main` directly (config
+defaults, the v11→v12 provider-schema normalizer, and the write-approval
+migration), not from docs or blog summaries, several of which turned out to
+be wrong or imprecise (noted below). Upstream has since shipped 0.18.1 and
+0.18.2 patch releases (2026-07-07); this review is against the 0.18.0
+"Judgment Release" behavior since that's what's pinned.
+
+### What's actually there at 0.18.0
+
+- **Memory is not three-tiered.** The built-in provider is two flat
+  Markdown files — `MEMORY.md` (agent notes, 2200-char cap) and `USER.md`
+  (user profile, 1375-char cap) — stored at `~/.hermes/memories/` and
+  injected into the system prompt. Defaults: `memory.memory_enabled: true`,
+  `memory.write_approval: false` (writes land with no review gate). A
+  `memory.provider` slot exists for 7 external providers (Honcho, Mem0,
+  Hindsight, Supermemory, OpenViking, holographic, retaindb) but is empty
+  by default — none of them is Engram, and there is no adapter hook for a
+  custom provider. The "semantic/working/episodic" framing in the survey is
+  secondary-source terminology describing the general memory-architecture
+  pattern, not a config surface this version exposes — there is no
+  independent on/off switch per tier.
+- **The skill-writing loop is present and on by default, unrestricted.**
+  `/learn` and the `skill_manage` tool distill reusable `SKILL.md` files
+  into `~/.hermes/skills/` after complex tasks, error-recovery, or user
+  correction. Governance is `skills.write_approval` (default `false` — free
+  writes) and `skills.guard_agent_created` (default `false` — no
+  content-pattern scan). One secondary source claimed a `skills.auto_create`
+  gate defaults it off; that key does not exist in source — treat that
+  claim as wrong.
+- **Provider-config schema is fine as pinned.** A closed 2026-04 upstream
+  issue (#8776) reported the `providers:` dict format being write-only
+  (migration wrote it, nothing read it back) — that's fixed on `main`:
+  `get_compatible_custom_providers()` now merges both `custom_providers:`
+  (legacy list) and `providers:` (dict) into one runtime view, and the
+  normalizer accepts `api`/`url`/`base_url` as URL-key aliases and both
+  list- and dict-shaped `models:`. Our `config/cli-config.local-endpoint.yaml`
+  preset (`providers.<key>.api`, `models:` as a string list) matches this
+  and needed no schema fix. A separate, still-**open** bug (#43713) affects
+  `hermes profile` sub-configs replacing rather than merging the parent's
+  `providers:` — irrelevant to us since this preset doesn't use profiles.
+
+### Engram boundary — the real question
+
+If hermes memory stays on with its default unrestricted writes, hermes and
+Engram independently accumulate the same class of fact (operator
+preferences, environment conventions) with no reconciliation between the
+two stores — not a corruption risk (they're separate files/DBs), but a
+drift risk: two silently-diverging "memories of the operator," one of which
+(hermes's) is invisible to session-bootstrap, Obsidian, and every other
+harness. Engram is this repo's designated cross-harness, cross-session
+memory system per the operator's global config. Lane assignment adopted
+here: **hermes's built-in memory tier goes off entirely**
+(`memory.memory_enabled: false` in the local-endpoint preset); Engram
+remains the sole long-term memory of record across all four harnesses.
+Skills are a different domain (hermes-local executable procedures, no
+Engram equivalent) — the self-improvement loop stays on, but
+`skills.write_approval: true` now gates agent-authored skill writes behind
+a review step (`/skills pending` / `/skills diff` / `/skills approve`),
+since they currently land with zero review and can carry inline shell
+(`skills.inline_shell`, off by default, unaffected by this change).
+
+Related: OPS-482 (in progress, separate branch) is defining
+checkpoint-vs-memory boundaries for shepherd's durable-patterns work — same
+category of problem, different subsystem. Not a blocker for this verdict;
+cross-check the two once OPS-482 lands.
+
+### Preset diff applied
+
+```diff
++ memory:
++   memory_enabled: false
++ skills:
++   write_approval: true
+```
+
+in `config/cli-config.local-endpoint.yaml`. `tests/test_presets.py` (YAML
+validity, localhost-only URLs, secretlessness) and
+`tests/test_pin_drift.py` (pin/README agreement) both still pass — this
+change doesn't touch the version pin.
+
+### Cluster-pin skew note
+
+unsigned-paas's kagent BYO agent is still on 0.16.0 against this package's
+0.18.0 — a two-minor-version gap. Tracked as a separate bump, not addressed
+here.

packages/harness-hermes/config/cli-config.local-endpoint.yaml

@@ -49,3 +49,22 @@ providers:
 # Alternate local endpoints (swap `api`):
 #   llama.cpp: http://localhost:8080/v1   (serve/serve-llamacpp.sh)
 #   vLLM:      http://localhost:8000/v1   (serve/serve-vllm.sh)
+
+# Memory/skill-loop boundary vs Engram (OPS-480, 2026-07-08; see README Verdict
+# section for the full reasoning). hermes's built-in memory store (two flat
+# files, ~/.hermes/memories/MEMORY.md + USER.md) would otherwise accumulate
+# the same class of cross-session facts about the operator as Engram, with no
+# reconciliation between the two. Engram is this repo's designated
+# cross-harness memory of record (session-bootstrap wiring, all four
+# harnesses) -- hermes's own tier stays off so there is exactly one canonical
+# long-term store.
+memory:
+  memory_enabled: false
+
+# The skill-writing self-improvement loop (/learn, skill_manage) lands
+# agent-authored SKILL.md files with no review gate by default upstream.
+# Skills are hermes-local procedural knowledge -- not Engram's domain -- so
+# the loop itself stays on, but agent-authored writes require a look before
+# they land.
+skills:
+  write_approval: true