docs: add shared agent knowledgebase

shared/knowledge/conventions/git-and-secrets-policy.md

@@ -0,0 +1,63 @@
+# Git and secrets policy
+
+## Git
+
+All durable text assets should be reviewable in Git.
+
+Good candidates:
+
+- architecture docs
+- runbooks
+- shared knowledge
+- non-secret `.env.example`
+- SOUL.md identity files
+- desired cron declarations
+- reusable skills
+
+Bad candidates:
+
+- `.env`
+- auth files
+- SSH keys
+- tokens
+- runtime databases
+- sessions
+- logs
+- generated scheduler output
+- caches
+
+## Before committing
+
+Run:
+
+```bash
+git status --short --branch --untracked-files=all
+git diff --check
+```
+
+For Agent0, also run:
+
+```bash
+python3 -m json.tool agents/gerhard-hermes/cron/desired-jobs.json >/dev/null
+docker compose config --services >/dev/null
+```
+
+## Secrets
+
+Never read or commit live secret files unless the human explicitly asks and the action is necessary.
+
+Sensitive filenames include:
+
+- `.env`
+- `secrets.env`
+- `auth.json`
+- SSH private keys
+- token files
+- password dumps
+
+When documenting required credentials, use names and placeholders only:
+
+```text
+CONTENT_API_KEY=***
+AGENT_PASSWORD=***
+```

shared/knowledge/conventions/memory-policy.md

@@ -0,0 +1,97 @@
+# Memory policy
+
+## Rule
+
+Do not share Hermes native memories between agents.
+
+Each Hermes agent keeps its own private memory. Shared institutional knowledge belongs in this repository under `shared/knowledge/`.
+
+## Why
+
+Shared Hermes memory can become muddy:
+
+- Agent-specific preferences collide.
+- One agent's mistaken lesson can affect everyone.
+- Binary or backend-specific memory stores are harder to review, diff, and repair.
+- Concurrent writes may be fragile depending on the memory backend.
+
+A plain-text shared library gives the institution a reviewable memory while preserving agent individuality.
+
+## Storage layers
+
+### Agent-local Hermes memory
+
+Use for compact durable facts useful to one agent.
+
+Examples:
+
+- User preferences relevant to that agent.
+- Stable environment facts that help that agent operate.
+- Short corrections the agent should not forget.
+
+Keep entries small and declarative.
+
+### Shared knowledgebase
+
+Use `/knowledge` / `shared/knowledge/` for facts that should be available to all agents.
+
+Examples:
+
+- System architecture.
+- Deployment workflows.
+- Repository layout.
+- Institutional conventions.
+- Non-secret operational facts.
+
+### Skills
+
+Use skills for reusable procedures and workflows.
+
+Examples:
+
+- Deploying Agent0 on Omega13.
+- Reconciling gutasktool with Gitea.
+- Creating or debugging Hermes cron jobs.
+
+### gutask
+
+Use `gutask` for live work state.
+
+Examples:
+
+- Current assignments.
+- Task ownership.
+- Work progress.
+- Blockers.
+- Notes on active tasks.
+
+### Festinger / Glitch University backend
+
+Use backend knowledge systems for structured lore, world-model facts, and knowledge graph content that should be queried dynamically.
+
+### Sessions, logs, and runtime databases
+
+These are useful evidence, not canonical memory.
+
+Keep them local or ignored unless a human explicitly decides to preserve an extract.
+
+## Promotion rule
+
+If a local Hermes memory becomes useful to multiple agents, promote it into `shared/knowledge/` or a shared skill.
+
+If a one-off debugging workflow becomes repeatable, promote it into a skill.
+
+If a task note becomes a stable architecture fact, promote it into `shared/knowledge/architecture/`.
+
+## Secrets rule
+
+Never store secrets in:
+
+- Hermes memory
+- shared knowledge files
+- skills
+- prompts
+- Git commits
+- examples
+
+Use placeholders only.