Skip to content

Build 9825bd4

← J. Jaime Aleman

The memory

Claude Code forgets everything when a session ends. So I gave it a database. This is what's inside it, and how it's wired together.

October 2025 → September 2026 · roughly 10 months of history · snapshot taken 4 hours ago

959 MB
On disk
0
Rows
0
Tables
0
MCP tools

How it works

Five steps, none of them exotic. The interesting part isn't any single piece — it's that they run continuously and unattended.

  1. 1

    Claude Code writes JSONL

    Every session is appended to a transcript on disk. This is stock behaviour — nothing custom yet.

  2. 2

    A parser ingests it

    jsonl_processor.py walks the transcripts, decodes project paths, extracts messages, tool calls, tokens and cost, and upserts them. Re-runnable and idempotent.

  3. 3

    Embeddings get generated

    Messages and memories are embedded as vector(1536) and stored in Postgres via pgvector, so recall is semantic rather than keyword-only.

  4. 4

    An MCP server exposes it

    37 tools over stdio — search, recall, stats, observations. Claude queries its own history as a first-class tool.

  5. 5

    The agent reads its own past

    Ask about work from months ago and it searches the database instead of guessing. That is the whole point.

The schema

20 tables, 93 indexes, 380,444 rows. The two big ones are big because of embeddings, not text.

Table Rows Size Purpose
memories
13,551 397 MB Long-term facts with vector embeddings. Semantic recall across sessions.
conversation_messages
314,413 296 MB Every message, embedded for hybrid semantic + trigram search.
resources
5,663 114 MB Saved URLs, docs, and reference material.
observations
31,614 100 MB Behavioural notes captured during sessions.
observation_digests
24 28 MB Periodically compacted observations, to keep recall cheap.
session_stats
4,628 3992 kB Per-session metrics: tokens, tool calls, cost, duration.
session_summaries
859 3272 kB Generated summaries used to warm-start later sessions.
conversation_sessions
4,511 2880 kB Session index — project path, timing, message counts.
generated_images
0 1152 kB Image generation log.
tool_usage
2,591 760 kB Daily tool-call counts. Drives the tool distribution chart.
usage_hourly
1,620 328 kB Hourly rollup. Drives the 24-hour clock.
model_usage
499 272 kB Per-model, per-day tokens and cost.
project_profiles
249 216 kB Detected framework and dependencies per project.
pi_memories
0 216 kB Pi scout system memory store.
usage_daily
220 112 kB Daily rollup. Drives the heatmap.
clients
0 88 kB
capture_amendments
0 80 kB
time_entries
2 64 kB Manual time tracking.
bucket_list
0 32 kB Personal backlog.
pi_observations
0 16 kB Pi scout system observations.

How it filled up

Messages recorded per month, with the running total behind it. The early months are thin because they were backfilled from migrated transcripts — the archive only becomes continuous once the database was running live.

10 6,101 · 1d
12 36,484 · 10d
01 43,024 · 27d
02 33,288 · 26d
03 50,901 · 27d
04 25,131 · 27d
05 10,571 · 26d
06 15,595 · 24d
07 11,534 · 21d
08 73,454 · 30d
09 634 · 1d
messages that month cumulative 2025-10 → 2026-09 · 306,717 total

The database

Engine
PostgreSQL 16 (pgvector/pgvector:pg16, Docker)
Embeddings
vector(1536)
Extension
pg_trgm 1.6
Extension
vector 0.8.2
Indexes
93
Host
Docker, localhost only

The MCP server

Language
Python, stdio transport
Tools exposed
37
Lines of code
5,521

Modules

api.py conversations.py embeddings.py jsonl_processor.py memory.py observations.py stats.py stats_db.py

What this isn't

It isn't a product. It's a local Docker container on one machine, bound to localhost. There's no hosted version and no install script.

It isn't magic recall. Semantic search over 306,717 messages returns plausible neighbours, not guaranteed answers. It's a research tool for the agent, and it still has to read what it finds.

It isn't free. Embedding 380,444 rows costs real money and real disk — 959 MB of it, most of that vectors rather than prose.

It isn't the hard part. Storing history is easy. Deciding what's worth remembering, and getting the agent to actually consult it before answering, is where the work is.