| name | llm-wiki-agent-memory-bridge |
| description | Decide what belongs in LLM-Wiki versus agent memory, AGENTS.md, CLAUDE.md, Auto Memory, Codex project memory, or other runtime memories. Use when instruction files are bloated, domain facts live in agent memory, or the user wants clean memory boundaries. |
| license | MIT |
| compatibility | Designed for Agent Skills-compatible coding agents. Requires read access to project instructions and wiki files; write access is optional for migration patches. |
| metadata | {"author":"po4yka","version":"1.0.0","install_scope":"self-contained"} |
LLM-Wiki Agent Memory Bridge
Goal
Separate durable domain knowledge from runtime/project memory so agents load the right context at the right time.
When to use
AGENTS.md/CLAUDE.md has grown large because domain facts were pasted into instructions instead of the wiki.
- A wiki page holds procedural agent prompts that belong in a Skill, not domain knowledge.
- Agent memory (Auto Memory, Codex project memory) duplicates or has drifted from existing wiki pages.
- You need to decide where a new piece of knowledge belongs: Skill, instruction file, or wiki.
- Onboarding a new agent (Claude Code, Codex, Cursor, OpenCode) to a project and need to map its existing memory surfaces first.
Inputs
AGENTS.md, CLAUDE.md, project memory files, agent settings.
- Existing wiki/vault.
- Target agents: Claude Code, Codex, Cursor, OpenCode, etc.
- Desired mode: audit, plan, or apply.
Procedure
1. Inventory memory surfaces
Find:
- instruction files;
- agent memory files;
- skills;
- wiki pages;
- query logs;
- project docs;
- hidden or generated memory directories.
2. Classify memory items
Use this routing:
| Content | Best home |
|---|
| stable procedures | Agent Skill |
| project conventions | AGENTS.md or CLAUDE.md pointer |
| domain facts | LLM-Wiki page with provenance |
| source-backed claims | wiki/sources, wiki/concepts, wiki/entities |
| user preferences | agent/user memory, not public wiki |
| temporary task state | current session or task log |
| sensitive facts | restricted wiki or local-only memory |
3. Detect bloat and drift
Flag:
- giant instruction files full of domain knowledge;
- skills containing volatile facts;
- agent memory duplicating wiki pages;
- wiki pages containing procedural prompts;
- missing pointers from instruction files to wiki index;
- volatile facts (pricing, release notes, maintenance status) embedded directly in a skill or instruction file instead of a wiki page whose claims get re-verified against fresh sources before reuse.
4. Propose moves
Create a migration table:
| Item | Current location | Proposed home | Reason | Risk |
|---|---|---|---|---|
5. Apply safe pointer pattern
Keep instruction files short. Replace long domain dumps with pointers such as:
For domain knowledge, read `wiki/index.md` first and then relevant linked pages.
Output
## Memory boundary summary
## Memory surfaces found
## Misplaced content
## Proposed moves
## Instruction-file pointer updates
## Risks
Safety gates
- Do not move user preferences into a shared wiki without approval.
- Do not delete memory content during migration; archive or link it.
- Do not hide domain facts in skills.
- Do not leave instruction files without pointers to the wiki.