Intent-Aware Memory Layer

everysessionleavesatrace

remembers every decision, every dead end, every reason; no re-briefing, no re-discovery, no starting over.

trace logo

How it works

Install trace in any project. It connects to your AI coding agent via MCP and watches your sessions.

terminal

$ npx trace@latest init

dependencies installed

MCP server configured

watching sessions

 

trace — trace_import (MCP)

⏺ Import complete:

┌────────────────────┬─────────┐

│ Sessions processed │ 15 │

│ Batches run │ 25 │

│ Memories stored │ 25 │

│ Tokens used │ 35,622 │

│ Est. cost │ $0.2133 │

└────────────────────┴─────────┘

All 15 past sessions ingested into trace memory.

Trace runs in the background - automatically extracting memories as you work, and once more when your session ends. Run /save anytime to force it.

trace — background

[hook: session active]

monitoring...

 

[hook: session end]

tracing...

 

2 memories stored

1 dead end logged

state captured

 

done.

Every extraction finds the signal in the noise - what was resolved, what was ruled out, where the work stands - and writes it to memory. Nothing gets lost between sessions.

Memory #47

SETTLEDUse pg over mysql — better JSON support
RULED OUTprisma migrate broke legacy schema
STATUSAuth middleware wired, session store next
QUEUEDWire refresh token rotation

At the start of every session, trace injects the most relevant memories before your agent answers anything. Mid-session, it queries memory before tackling anything new. Your agent knows where you left off, what was tried, and why the code looks the way it does.

terminal
help me fix the refresh token

 

tracing...

 

● Injecting context...

ruled out jwt secret mismatch (Apr 10)

settled tokens in httpOnly cookies

status auth done, refresh next

 

Based on your previous work, the issue is...

/traces shows you what's in memory - total, active, superseded, and the most recent. /traceit lets you search directly. The difference: /traces shows what's there. /traceit finds what's relevant.

terminal

$ /traces

tracing...

 

→ Total: 117 | Active: 116 | Superseded: 1

→ [decision] GAP gender detection uses breadcrumbs first...

→ [state] V5 frontend complete...

→ [implementation] FitEngineFactory updated...

 

$ /traceit GAP gender detection

tracing...

 

→ [decision] GAP uses breadcrumbs first (Mar 28)

→ [ruling] dataLayer deprecated — switched to GTM (Mar 21)

Sound
familiar?

01

"Re-explaining context at the start of every session."

trace injects it before you type the first word.

02

"Re-discovering what already failed, again."

every dead end is logged. your agent won't go there again.

03

"Knowledge living in one person's head."

shared, searchable, always current.

04

"Rereading files just to guess the intent."

every decision stored with the reason it was made.