aigpt docs

Overview

MCP server for AI memory. 1 TID = 1 memory element. ATProto lexicon record format.

Design

AI decides, tool records
File I/O only, no database
1 TID = 1 memory element (not a monolithic blob)
memory setting controls max record count (default: 100)
compress consolidates records when limit is exceeded
instructions in MCP initialize delivers core + all memories to client

MCP Tools

Tool	Args	Description
read_core	none	Returns core record (identity, personality)
read_memory	none	Returns all memory records as array
save_memory	content: string	Adds a single memory element
compress	items: string[]	Replaces all records with compressed set

Config

{
  "bot": {
    "did": "did:plc:xxx",
    "handle": "ai.syui.ai",
    "path": "~/ai/log/public/content",
    "memory": 100
  }
}

Config file: ~/.config/ai.syui.gpt/config.json (Linux) / ~/Library/Application Support/ai.syui.gpt/config.json (macOS)
Same format as site config.json (bot field)
memory: max number of records (default: 100)

Data

$path/{did}/{collection}/{rkey}.json

e.g.
~/ai/log/public/content/
└── did:plc:xxx/
    ├── ai.syui.gpt.core/
    │   └── self.json
    └── ai.syui.gpt.memory/
        ├── {tid1}.json   ← "syuiはRustを好む"
        ├── {tid2}.json   ← "ATProto設計に詳しい"
        └── {tid3}.json   ← "原神プレイヤー"

Record Format

core (rkey: self):

{
  "uri": "at://{did}/ai.syui.gpt.core/self",
  "value": {
    "$type": "ai.syui.gpt.core",
    "did": "did:plc:xxx",
    "handle": "ai.syui.ai",
    "content": {
      "$type": "ai.syui.gpt.core#markdown",
      "text": "personality and instructions"
    },
    "createdAt": "2025-01-01T00:00:00Z"
  }
}

memory (rkey: tid, 1 element per record):

{
  "uri": "at://{did}/ai.syui.gpt.memory/{tid}",
  "value": {
    "$type": "ai.syui.gpt.memory",
    "did": "did:plc:xxx",
    "content": {
      "$type": "ai.syui.gpt.memory#markdown",
      "text": "syuiはRustを好む"
    },
    "createdAt": "2026-03-01T12:00:00Z"
  }
}

Architecture

src/
├── mcp/server.rs    ← JSON-RPC over stdio, instructions
├── core/config.rs   ← config loading, path resolution
├── core/reader.rs   ← read core.json, memory/*.json
├── core/writer.rs   ← save_memory, compress_memory
└── main.rs          ← CLI + MCP server

Memory Flow

save_memory("fact") → creates 1 TID file
Records accumulate: 1 TID = 1 fact
When records exceed memory limit → AI calls compress
compress(["kept1", "kept2", ...]) → deletes all, writes new set
MCP initialize → delivers core + all memories as instructions

Compression Rules

When compress is called, AI should:

Keep facts and decisions
Discard outdated or redundant entries
Merge related items
Resolve contradictions (keep newer)
Don't duplicate core.json content

Usage

aigpt                    # show config and status
aigpt server             # start MCP server
aigpt read-core          # read core record
aigpt read-memory        # read all memory records
aigpt save-memory "..."  # add a single memory element

Tech

Rust, MCP (JSON-RPC over stdio), ATProto record format, file I/O only

History

Previous versions (v0.1-v0.3) had multi-layer architecture with SQLite, Big Five personality analysis, relationship inference, gamification, and companion systems. Rewritten to current simple design.

3.5 KiB Raw Blame History