Architecture

System Layers

┌─────────────────────────────────┐   ┌──────────────────────────────┐
│      Agent Tools                │   │      Notebook Skill          │
│  memory_search, block_update    │   │  ingest / lint / query-file  │
├─────────────────────────────────┤   ├──────────────────────────────┤
│      MemoryManager              │   │      WikiManager             │
│  Orchestration & extraction     │   │  Bootstrap schema/index/log  │
├─────────────┬───────────────────┤   ├──────────────────────────────┤
│  LanceDB    │  MarkdownStore    │   │  Obsidian Vault              │
│  (search)   │  (source of truth)│   │  /mnt/notebook               │
├─────────────┴───────────────────┤   └──────────────────────────────┘
│      Session Layer              │
│  Transcripts, state, lifecycle  │
└─────────────────────────────────┘
       Conversation Memory                   LLM Wiki

Storage Tiers

Markdown Files (/shared/memory/)

Human-readable source of truth. Agent-facing — the agent can read/write these files directly via ReadFileTool/WriteFileTool.

• Daily logs (YYYY-MM-DD.md): Append-only, timestamped facts per conversation turn
• MEMORY.md: Curated long-term summary, auto-updated by refresh_core_memory_facts()

LanceDB (.suzent/memory/)

Vector search index. Rebuilt from markdown if corrupted via MarkdownIndexer.reindex_from_markdown().

• Core memory blocks (persona, user, facts, context)
• Archival memories with vector embeddings
• Hybrid search: semantic + full-text + importance + recency

Session Data (.suzent/)

Internal operational data. Not agent-facing — accessed via API endpoints.

• Transcripts (.suzent/transcripts/{id}.jsonl): Append-only per-session conversation logs
• State snapshots (.suzent/state/{id}.json): Inspectable JSON agent state mirrors

Components

MarkdownMemoryStore (markdown_store.py)

Human-readable persistence layer

• Manages files in /shared/memory/ (cross-session, agent-accessible)
• Daily log format: lean single-line-per-fact (- [category] content \tag1 tag2``)
• Methods: append_daily_log(), get_recent_logs(), write_memory_file(), read_memory_file()

LanceDBMemoryStore (lancedb_store.py)

Search index layer

• LanceDB connection management
• Core memory block CRUD with scoping priority (chat > user > global)
• Archival memory with vector embeddings
• Semantic and hybrid search

MemoryManager (manager.py)

Orchestration layer

• Dual-write: every extracted fact goes to both LanceDB and markdown
• Automatic LLM-based fact extraction (one concise sentence per fact)
• Deduplication via similarity threshold (0.85)
• Core memory refresh when high-importance facts are found
• Transcript linkage: facts track source_session_id, source_transcript_line, source_timestamp

WikiManager (wiki_manager.py)

LLM Wiki bootstrap

• Initializes the Obsidian-style vault at the notebook root (/mnt/notebook in sandbox)
• Seeds three navigation files on first init: schema.md (from skills/notebook/schema_example.md), index.md, log.md
• After bootstrap, the agent owns the vault entirely — WikiManager makes no further writes
• Separate from conversation memory: no LanceDB involvement, no automatic extraction

MarkdownIndexer (indexer.py)

Recovery mechanism

• reindex_from_markdown(): Rebuilds LanceDB from markdown daily logs
• Parses facts from markdown format, generates embeddings, stores in LanceDB

TranscriptIndexer (indexer.py)

Cross-session search

• Chunks JSONL transcripts (~400 tokens, 80 overlap) into LanceDB
• Enables semantic search across past session conversations
• Opt-in via transcript_indexing_enabled config

ContextCompressor (context_compressor.py)

Context window management with pre-compaction flush

• Before compressing steps, extracts facts from steps about to be removed
• Feeds synthetic ConversationTurn to MemoryManager for dual-write
• Ensures no valuable context is lost when the context window is trimmed

Memory Tools (tools.py)

Agent interface

• MemorySearchTool: Semantic search across archival memory
• MemoryBlockUpdateTool: Update core blocks (replace, append, search_replace)
• Thread-safe via asyncio.run_coroutine_threadsafe()

Memory Context (memory_context.py)

Prompt templates

• FACT_EXTRACTION_SYSTEM_PROMPT: One concise sentence per fact, no filler
• CORE_MEMORY_SUMMARIZATION_PROMPT: Max 200 words, bullet points only
• format_core_memory_section(): Includes Memory Workspace location (/shared/memory/)
• format_retrieved_memories_section(): Single-line format per memory

Session Components

TranscriptManager (session/transcript.py)

• Append-only JSONL files at .suzent/transcripts/{session_id}.jsonl
• Each line: {"ts", "role", "content", "actions", "meta"}
• Thread-safe with async locks per session

StateMirror (session/state_mirror.py)

• Writes inspectable JSON snapshots to .suzent/state/{session_id}.json
• Parses JSON v2 agent state; writes placeholder for legacy pickle

SessionLifecycle (session/lifecycle.py)

• SessionPolicy: daily reset hour, idle timeout, max turns
• should_reset(): checks all policies against session metadata

Agent Serializer (core/agent_serializer.py)

• JSON v2 format: Human-readable, inspectable agent state
• Backward-compatible: auto-detects and loads legacy pickle format
• Serializes smolagents step objects (ActionStep, PlanningStep, TaskStep, FinalAnswerStep)

Data Flow

Write Path (Extraction + Dual-Write)

Conversation Turn (User + Agent + Response)
  ↓
manager.process_conversation_turn_for_memories()
  ↓
LLM extracts concise facts
  ↓
┌────────────────┬─────────────────────┐
│ LanceDB store  │ Markdown daily log  │
│ (search index) │ (source of truth)   │
└────────────────┴─────────────────────┘
  ↓
If High Importance → refresh_core_memory_facts() → MEMORY.md

Read Path (Memory Injection)

User Query
  ↓
manager.retrieve_relevant_memories()
  ↓
Generate embedding → Hybrid search (LanceDB)
  ↓
Format results (single-line per memory)
  ↓
Inject into agent prompt

Pre-Compaction Flush

Context compression triggered
  ↓
_pre_compaction_flush(steps_to_compress)
  ↓
Build synthetic ConversationTurn from steps
  ↓
Feed to memory extraction pipeline (dual-write)
  ↓
Compress steps into summary

Session Lifecycle

New message arrives
  ↓
Write JSONL transcript entry
  ↓
Extract memories (dual-write)
  ↓
Check compression → pre-compaction flush if needed
  ↓
Persist state (JSON v2 + StateMirror)
  ↓
Update last_active_at, turn_count

Memory Scoping

User-Level

• Scope: All chats
• Storage: user_id="x", chat_id=NULL
• Use: Preferences, facts, persona

Chat-Level

• Scope: Single conversation
• Storage: user_id="x", chat_id="y"
• Use: Current context, session state

Global

• Scope: All users/chats
• Storage: user_id=NULL, chat_id=NULL
• Use: Default persona

Priority

1. Chat-specific (most specific)
2. User-level (persistent)
3. Global (fallback)

File Structure

src/suzent/memory/
├── __init__.py          # Exports
├── lancedb_store.py     # LanceDB search index
├── markdown_store.py    # Markdown source of truth (/shared/memory/)
├── indexer.py           # MarkdownIndexer + TranscriptIndexer
├── manager.py           # Orchestration & dual-write
├── wiki_manager.py      # LLM Wiki bootstrap (schema/index/log)
├── memory_context.py    # Prompt templates
├── tools.py             # Agent tools
├── models.py            # Pydantic models
└── lifecycle.py         # Initialization

skills/notebook/
├── SKILL.md             # Notebook skill definition
├── schema_example.md    # Default schema seeded into new vaults
├── ingest.md            # Ingest procedure
└── lint.md              # Lint procedure

src/suzent/session/
├── __init__.py
├── transcript.py        # JSONL transcripts
├── state_mirror.py      # Agent state snapshots
└── lifecycle.py         # Session reset policies

src/suzent/core/
├── agent_serializer.py  # JSON v2 agent state
├── context_compressor.py # Pre-compaction flush
└── chat_processor.py    # Integration point

src/suzent/routes/
└── session_routes.py    # Transcript/state/memory APIs

Data Layout

data/sandbox-data/
  shared/                           # Cross-session workspace (agent sees /shared/)
    memory/
      MEMORY.md                     # Curated long-term memory
      2026-02-08.md                 # Daily append-only logs

/mnt/notebook/  (or $MOUNT_NOTEBOOK)  # LLM Wiki vault (Obsidian-style)
  schema.md                         # Vault conventions — agent reads first
  index.md                          # Catalog of synthesized pages
  log.md                            # Append-only operation log
  <user notes and folders...>

.suzent/
  chats.db                          # SQLite + session lifecycle fields
  memory/                           # LanceDB search index
  transcripts/
    {session_id}.jsonl              # Per-session conversation logs
  state/
    {session_id}.json               # Agent state snapshots

Design Principles

1. Markdown as Source of Truth - Human-readable, git-friendly, recoverable
2. Dual-Write - Every fact persisted to both markdown and LanceDB
3. No Context Loss - Pre-compaction flush captures facts before compression
4. Inspectable State - JSON v2 replaces opaque pickle; state mirrors on disk
5. Separation of Concerns - Agent-facing (/shared/) vs internal (.suzent/)
6. Async by Default - Non-blocking I/O throughout
7. Backward Compatible - Legacy pickle deserialization, nullable migrations
8. Wiki as Agent-Owned Knowledge - The LLM Wiki vault is bootstrapped once, then entirely maintained by the agent via direct file tool access; the system does not auto-write to it. Pattern from Karpathy's LLM Wiki: raw sources → LLM-maintained wiki → schema conventions