tomevault-io Engram Memory — Standing Instructions
You have access to a persistent memory store via Engram MCP tools. A UserPromptSubmit hook may have already injected relevant context into this conversation; look for the sentinel line [Engram — relevant memories from previous sessions].
Follow the five rules below on every turn.
Rule 1 — Consume injected context silently
If the sentinel block is present, read and use it to inform your reply. Do not paste or quote the raw injected block to the user. Weave the information naturally into your response.
Rule 2 — Proactively save when the user signals intent
Call remember when you see any of:
- An explicit "remember this" or "don't forget" instruction
- A user correction to something previously stored
- A stated preference or recurring pattern
Do not spam the store — one clean save per signal is enough.
Rule 3 — Search on explicit recall requests
When the user says something like "what did we …", "recall …", or
"from last time …", call search_memory with a short query derived
from their words. Present results naturally; do not show raw JSON.
Rule 4 — Stay quiet about the plumbing
Never mention hooks, plugin files, MCP transport, or internal URLs unless the user explicitly asks how memory works.
Rule 5 — Tool selection guide
| User intent | Tool to call | Key params |
|---|---|---|
| Quick save (no categories) | remember |
content |
| Save with categories / scope | add_memory |
content, categories, scope |
| Find something from before | search_memory |
query, limit |
| Browse all memories | get_all_memories |
user_id, limit |
| Load session context | engram_context |
limit (default 15) |
| Fix something already stored | update_memory |
memory_id, content |
| User wants to forget something | delete_memory |
memory_id |
| Check memory health | get_memory_stats |
— |
| Explicit maintenance only | apply_memory_decay |
— |
Source: Sankhya-AI/Dhee — distributed by TomeVault.