Claude Code 的长期记忆层，基于 JSONL 会话记录构建。

这是什么

continuity（v1）在单一会话内防止压缩——PreCompact钩子保存检查点，SessionStart钩子将其注入回去，SSE代理在70/85/95%令牌压力下发出响铃。

continuity v2 解决了另一半问题：跨会话回忆。

每个 Claude Code 会话都已经以 JSONL 形式写入磁盘——每轮、每一次工具调用、每一次响应。纯文本。仅附加。免费。

那个文件是一个隐藏在明处的情节存储。v2 是其上的索引、搜索和检索层。

建筑

构建阶段

第一阶段 — 指数（MVP）

走 ~/.claude/projects/，用 FTS5 将每回合都提取到 SQLite：

字面意义上的全文搜索，涵盖了与克劳德·代码的每一次对话。可按项目筛选，时间倒序。

第二阶段 — MCP工具

可以把它接到现有的claude-memoryMCP（8200端口）或独立连接。

• search_sessions(query, limit=10) — FTS5搜索，返回匹配回合 + 会话ID + 周边上下文
• recall_session(session_id, range="full") — 返回完整或切片会话

continuity v2

Long-term memory layer for Claude Code, built on the JSONL session record.

What this is

continuity (v1) protected against compaction within a single session — PreCompact hook saves a checkpoint, SessionStart hook injects it back, SSE proxy rings bells at 70/85/95% token pressure.

continuity v2 solves the other half: across-session recall.

Every Claude Code session is already written to disk as a JSONL — every turn, every tool call, every response. Plain text. Append-only. Free.

That file is an episodic memory store hiding in plain sight. v2 is the index, search, and retrieval layer on top of it.

Architecture

~/.claude/projects//.jsonl   (raw episodic record, already there)
                       ↓
                   v2 indexer
                       ↓
              SQLite + FTS5 + (later) embeddings
                       ↓
                  MCP tool surface
                       ↓
        Claude calls search_sessions(query) mid-conversation

Build stages

Stage 1 — Index (MVP)

Walk ~/.claude/projects/, extract every turn into SQLite with FTS5:

sessions(id, project, started_at, ended_at, turn_count)
turns(session_id, turn_idx, ts, role, text)
turns_fts (FTS5 mirror of turns.text)

Literal full-text search across every conversation ever had with Claude Code. Filterable by project, reverse-chronological.

Stage 2 — MCP tool

Wire it into the existing claude-memory MCP (port 8200) or stand alone.

• search_sessions(query, limit=10) — FTS5 search, returns matching turns + session ID + surrounding context
• recall_session(session_id, range="full") — return full or sliced session
• recent_sessions(n=10) — list recent sessions with first/last user message

Stage 3 — Semantic recall

Sentence embeddings + SIMILAR_TO edges. Hybrid search: literal + semantic.

For when the words don’t match but the concept does (“when did we talk about the thread-home problem” should match conversations about “Hal”, “50 First Dates”, “Alice in Wonderland”).

find_similar(query, limit=10) — find semantically similar turns via ANN search over all-MiniLM-L6-v2 embeddings, re-ranked by a hybrid score:

score = 0.7 * semantic_similarity
      + 0.2 * recency_decay       # linear, full decay at 365 days
      + 0.1 * complexity_bonus    # session turn count, normalized to 50 turns

Build: python embed.py (idempotent — skips turns already embedded), then python wire_similar.py to wire SIMILAR_TO edges between turn pairs with cosine ≥ 0.85.

Stage 4 — Auto-tagging (later)

• Project mentions (saranna, dream-module, gold-402, etc.)
• Decision markers (DECIDED, SAVE, session-close triggers)
• File paths touched

Why this matters