DDDFXYqiming/Agent_Extensions

Memory Management (DSH Edition)

Cross-session long-term memory: L1 index injection (visible every turn) + L2 environment facts + L3 task experience.

Trigger Timing

Reading (When to Look Up Memory)

• At the start of a new task: if the task involves the local environment, tool configuration, specific tech stacks, or similar things done before → first memory_list to see what's available, then memory_read to retrieve relevant entries
• When encountering difficulties / hitting obstacles: memory_read to check if there is a relevant SOP (keyword match against sops/ filenames or facts section)
• Memory index in model prompt (memory:index): L1 existence index visible every turn — when related trigger words are seen, actively call memory_read/memory_list to get details

Writing (When to Consolidate Memory)

When a task is completed (or a phase is completed) and there is information validated by successful actions, call memory_write:

Can Write (must include evidence):

• Environment-specific facts: paths, configurations, measured parameters, tool behavior (→ entry_type: fact)
• Complex task experience: pain points that required multiple retries to succeed, hidden preconditions, stable steps (→ entry_type: sop)
• General hard rules (→ also possible via memory_write sop or directly suggest maintaining [RULES])

Forbidden to Write (writing would be pollution):

• ❌ Information without verification evidence (no action, no memory)
• ❌ Model's inherent knowledge, reasoned guesses, unverified assumptions
• ❌ Ephemeral state: timestamps, PIDs, temporary paths, one-time IDs
• ❌ General common sense, log records, details of reasoning steps

Index Synchronization

• memory_write automatically syncs the L1 index; manual operation is usually unnecessary
• After manually modifying facts.md / sops/ files → memory_index to rebuild the index
• When the index exceeds 30 lines: trim [RULES] (word-level edits, do NOT overwrite the full text)

Storage Layout

<home>/.dsh/memory/
├── memory_management_sop.md   L0 meta-rules (how to manage memory)
├── index.txt                  L1 index (≤30 lines: L2/L3 existence list + RULES)
├── facts.md                   L2 environment facts (## SECTION)
├── sops/*.md                  L3 task experience
└── file_access_stats.json     read frequency statistics

Tools

Tool	Purpose
memory_list	List all memory (facts + sops + index line count)
memory_read	Read specified memory (index / fact topic / sop filename)
memory_write	Write memory (fact/sop, evidence required)
memory_index	Rebuild the L1 index automatic section

Principles

1. Action Verification: No Execution, No Memory. Only write information that has been successfully verified.
2. Minimum Sufficiency: Keep content as short as possible; only record information whose forgetting would lead to costly retries.
3. Do Not Delete or Alter Verified Facts: You may compress or migrate, but never discard.
4. Active Writing: Memory writes are always initiated by the model/user; the system will never automatically modify configurations or install plugins.