Xanther Memory Engine
Persistent, cross-session memory for coding agents. XME remembers the decisions you made, the approaches you rejected, and the conventions you follow — so your agent never re-learns your codebase or repeats a mistake it already made.
Your agent forgets everything, every session
Every new chat starts from zero. Your agent doesn't remember what it tried yesterday, which approaches failed, or why you structured things the way you did. So it re-reads your code, re-proposes rejected ideas, and re-explains your stack — burning tokens and your time. On a project that spans weeks, that reset happens dozens of times. XME ends it.
A day (and month) in the life
Memory that compounds over time
Real coding work spans days, sessions, and teammates. Here's how cross-session memory changes each moment.
Monday: you start a migration
Without memory: You ask your agent to migrate the payments module to the new API. It explores the code, picks an approach, and you stop for the day.
With XME: XME captures the decision ("migrate via adapter pattern"), the files explored, and the rejected approach ("direct rewrite — too risky").
Thursday: you come back
Without memory: Three days later you reopen the task in a fresh session. A stateless agent starts from zero — re-reads everything, re-proposes the rewrite you already rejected.
With XME: With XME, the agent recalls: "We chose the adapter pattern Monday, rejected the direct rewrite. Here is what is left." It resumes instantly.
Next sprint: a teammate joins
Without memory: A new engineer picks up the module. Their agent has none of the context — the "why" lives in a Slack thread nobody can find.
With XME: Their agent inherits the shared project memory: decisions, conventions, and the full investigation trail. Onboarding drops from days to minutes.
A month later: a bug resurfaces
Without memory: The same race condition you fixed weeks ago comes back. The agent has no memory of the original diagnosis.
With XME: XME surfaces the past episode: "Fixed this race in send() on Aug 20 — root cause was missing lock." The fix is immediate, not a re-investigation.
How it works
Three layers of memory
XME stores memory the way a good engineer does — verbatim history, distilled facts, and live working state.
Verbatim session transcripts
Every conversation, tool call, and file touched — stored so the agent can replay exactly what happened weeks ago.
Decisions · Attempts · Preferences
LLM-extracted structured facts: "chose JWT over sessions", "tried Redis caching — reverted, too complex", "team uses snake_case".
Live state per project + user
The current task state, open threads, and what is left to do — UPSERTed so the agent resumes mid-flow.
Why it matters
Built for long-horizon projects
Long-horizon projects
Multi-week features span dozens of sessions. XME keeps the thread — what was decided, what failed, what remains — so progress compounds instead of resetting every conversation.
Decision audit trail
Every architectural choice is logged with its rationale and timestamp. When someone asks "why did we do it this way?" six months later, the answer is one query away.
No repeated mistakes
Failed approaches are remembered as facts. The agent will not re-propose the caching layer you already tried and reverted — it learns from the project history, not just the current file.
Team knowledge that persists
Memory is a file in your repo (or a shared server). It outlives any single developer, chat, or IDE. The knowledge stays with the codebase.
Capabilities
Persistent Memory
Your agent remembers past interactions, decisions, and context — even across restarts and days.
- ▸Episodic session store
- ▸Semantic + full-text recall
- ▸Automatic deduplication
Cross-Session State
State carries over between sessions so your agent picks up right where it left off — no re-explaining.
- ▸Session continuity
- ▸Working context UPSERT
- ▸Per-project + per-user scoping
Decision Tracking
Every architectural decision, rejected approach, and rationale is captured as a queryable fact.
- ▸Decision log with rationale
- ▸Failed-attempt memory
- ▸Reasoning traces
Preferences & Conventions
Learns your coding style, naming conventions, and architectural preferences over time.
- ▸Style learning
- ▸Convention enforcement
- ▸Adaptive suggestions
Get started in 30 seconds
Install once. Memory captures automatically.
XME installs IDE hooks that record every session — no manual note-taking. Works with Kiro, Claude Code, Cursor, and any MCP-compatible tool.
pip install xanther-xme xme hook install . # auto-captures every session xme start my-project # memory starts now # Later — recall anything: xme search my-project "why did we choose the adapter pattern?"