Memory Gregory owns across LLMs
Created
Memory Gregory owns across LLMs
Researched 2026-10-07 (Claude Opus,
read-only, Astra-reviewed); report in <path>.
Answer
Not a new memory server, not now. Two layers he already owns cover it:
| Layer | Holds | Who uses it | Why |
|---|---|---|---|
| This wiki + AGENTS.md / CLAUDE.md | How his system works, decisions, research, his norms | Claude, Codex, Serenity (reads), Gregory (Obsidian) | Plain Markdown in git on his NAS, already read by every runtime, has a writer protocol and backups |
| hindsight | Serenity’s personal, conversational memory | Serenity only | Self-hosted on the assistant’s guest, MIT, speaks MCP; kept away from coding agents so his life memory doesn’t spread to more vendors |
The silo was Claude’s per-project memory folders; the fix was a rule plus a sweep, done the same day: durable facts go to the wiki (how gregory works with agents), tool memory keeps pointers only.
Options compared
| Option | Verdict |
|---|---|
| A. Wiki + files (tightened) | Picked |
| B. Hindsight as the shared layer for Claude/Codex | Keep for Serenity only (privacy: its extraction LLM is OpenAI) |
| C. Basic Memory (MCP over Markdown, local index) | Revisit if agents need search the index and grep can’t give |
D. MCP reference server-memory |
Dropped: basic, substring search only |
| E. Graphiti (Zep) | Dropped: heavy (graph DB + LLM extraction) |
| F. Mem0 / OpenMemory | Dropped: no longer local |
| G. Letta | Dropped: a whole agent runtime, would replace Hermes |
What would change the answer
A second assistant that needs his personal context, a local extraction model for Hindsight, or a wiki past ~1,000 pages where grep stops being enough.