首页 › concepts › deepseek tui memory
deepseektuimemorypersistencellmprompt-engineering type: concept 创建: 2026-05-04 更新: 2026-05-04

DeepSeek-TUI User Memory

A persistent cross-session note file injected into the system prompt every turn. Gives the model durable memory of user preferences and project conventions without repeating them per session.

Overview

Memory is opt-in. When disabled (default), nothing is loaded and the remember tool is not surfaced. Zero overhead for users who haven't asked for it.

When enabled: ~/.deepseek/memory.md is read at every prompt-build call and injected verbatim into the system prompt as an XML block above the volatile-content boundary — inside DeepSeek's prefix cache so it survives turn-over-turn.

Enabling

# Via environment
export DEEPSEEK_MEMORY=on

# Via config (~/.deepseek/config.toml)
[memory]
enabled = true

Accepted truthy values: 1, on, true, yes, y, enabled.

Restart TUI after toggling. Memory file path can be overridden:

memory_path = "~/.deepseek/memory.md"  # default

Or via DEEPSEEK_MEMORY_PATH env var (wins over config file).

Three Ways to Add Notes

1. # Composer Prefix (No Turn Fires)

Type a line starting with # (but not ## or #!) in the composer:

# remember that this repo prefers 4-space indentation

The TUI intercepts the input, appends a timestamped bullet to the memory file, and confirms in the status line. No turn fires — you can keep typing your real question.

2. /memory Slash Command

Subcommand Effect
/memory or /memory show Show resolved path and current contents inline
/memory path Print just the resolved path
/memory clear Replace file with empty marker
/memory edit Print ${VISUAL:-${EDITOR:-vi}} <path> to open in editor
/memory help Show command-specific help and current path

3. remember Tool (Model-Initiated)

When memory is enabled, the model gets an auto-approved remember tool:

{
  "name": "remember",
  "description": "Append a durable note to the user memory file...",
  "input_schema": {
    "type": "object",
    "properties": {
      "note": { "type": "string" }
    },
    "required": ["note"]
  }
}

Auto-approved because writes are scoped to the user's own memory file — gating would defeat the purpose of automatic memory capture.

The tool description explicitly tells the model: durable, single-sentence notes only. Transient task state ("I'm currently editing foo.rs") is harmless but wasteful.

System Prompt Injection Format

<user_memory source="/Users/you/.deepseek/memory.md">
- (2026-05-03 22:14 UTC) prefer pytest over unittest
- (2026-05-03 22:31 UTC) this codebase uses 4-space indentation
</user_memory>

The block sits above the volatile-content boundary so it stays inside DeepSeek's prefix cache.

File Format

Plain Markdown with timestamped bullets:

- (2026-05-03 22:14 UTC) prefer pytest over unittest
- (2026-05-03 22:31 UTC) this codebase uses 4-space indentation
- (2026-05-04 09:02 UTC) all PRs need 2 reviewers before merge

The loader reads the whole file verbatim. Timestamps are convention (not enforced) for human grooming. Files over 100 KiB are loaded but truncated with a marker.

What Stays Out

Category Why Not
Secrets (API keys, tokens) Plain text on disk, injected verbatim into system prompt
Transient task state Changes every session; doesn't belong in cross-session memory
Conversation snippets Use the note tool instead
Long instructions Use AGENTS.md (project-level) or skills (reusable packs)

Memory vs Project Instructions

Memory is user-scoped, not repo-scoped:

  • Memory → personal preferences that follow you across repos and sessions
  • Project instructions (AGENTS.md, .deepseek/instructions.md) → repo-specific conventions that travel with the codebase

The memory loader currently reads one resolved file path. @path imports are not supported.

Configuration Reference

Setting Default Override
Memory enabled false [memory] enabled = true or DEEPSEEK_MEMORY=on
Memory file path ~/.deepseek/memory.md memory_path = "..." or DEEPSEEK_MEMORY_PATH=
Max file size 100 KiB (none; truncation marker shown)

Sub-Agent Behavior

Sub-agents inherit the user's memory file and can also use the remember tool. This means a sub-agent can capture a preference it discovers mid-task, and the parent session (and future sessions) will also know it.

Privacy

  • File lives entirely on your machine in ~/.deepseek/
  • Never uploaded to any cloud service
  • Included inline in the system prompt sent to the LLM provider
  • Provider-agnostic: switching providers uses the same file
  • Per-user, not per-project

Related