engram@core:~$ cat ./about.md
memory core · v0.1 · operational

the memory your engineering org keeps forgetting

engram is an engineering intelligence platform. it ingests your github repos and notion workspaces, cross-validates docs against code, and answers grounded in the verified org knowledge that actually shipped. no hallucinated tribal lore.

stackretrievaldeploynext · fastapi · pgvectorclaude sonnet · voyage codevercel · fly.io
engram@core:~/workspace$ ask
youwhy does our deploy fail on cold-start in eu-west?
explaintop-4 retrievedverified
>cold-start in eu-west-1 times out because the celery worker pulls the embedding model on first boot [1]. the deploy script doesn't warm the cache before health-check[2]. the runbook in notion calls for a 90s grace period; current k8s manifest sets 30s [3]— that's the mismatch.
# k8s/eu-west.yaml — patch grace period
readinessProbe:
  httpGet: { path: /healthz, port: 8000 }
  initialDelaySeconds: 90   # was 30
  periodSeconds: 10
retrievalk=4
├─01git backend/celery_app.pyrel ████████░░ 84%
├─02git scripts/deploy.shrel ███████░░░ 71%
├─03ntn runbook · cold-start eurel ██████░░░░ 66%
└─04git k8s/eu-west.yamlrel █████░░░░░ 52%
ONLINE4 sources1.2s · 720 tok↩ enter to send
indexed objects
2.4M
across customer workspaces
p50 retrieval
180ms
k=10 · pgvector + rerank
staleness caught
34%
of notion pages disagree with code
cold-start to answer
< 6s
from repo connect to first chat
// 01 ── modules

seven surfaces, one memory.

every module reads from the same indexed core. retrieval, ranking and cross-validation are shared.
01 · ASKchat

workspace chat

ask in natural language, get answers cited to specific files and notion pages. intent detection routes between explain and generate paths.

  • streaming sse with retrieval inline
  • per-message intent tags
  • conversation history, follow-ups
02 · SRCingest

connections

github oauth + notion oauth. periodic re-index via celery. one-click connect, progress in-line.

  • repo and workspace level scoping
  • incremental re-embedding on push
  • ingest-job failures surfaced as toasts
03 · LIBstore

library

browse indexed sources, inspect chunk boundaries, drop pages that shouldn't be retrieved.

  • file tree + chunk preview
  • exclude rules per resource
  • last-indexed timestamps
04 · JOBqueue

jobs

watch celery queues, inspect failed re-indexes, retry with one click. operator-grade observability.

  • per-job structured logs
  • sentry-style error summaries
  • retry with original parameters
05 · OUTsave

outputs

save any answer as a durable artifact — code snippet, summary, or report. shareable, citeable, versioned.

  • save-as: code · summary · report
  • permalink with source pins
  • type filter + pagination
06 · DIFverify

compare

cross-validate notion docs against the code they describe. flag stale runbooks before they cost an oncall.

  • doc-vs-code diff with citations
  • staleness scoring per page
  • code is the source of truth
07 · CFGadmin

settings

auth, integrations, model preferences, mcp tokens. per-conversation provider overrides on the roadmap.

  • github oauth, notion oauth
  • 30-day mcp jwt for power users
  • byo api key (v1.5)
+ EXTmcp

mcp plugin

same retrieval, same answers — exposed via the model context protocol. plug engram into claude desktop, cursor, zed.

  • paste-token install (30-day jwt)
  • shared retrieval contract
  • tool-call surface mirrors web ui
// 02 ── flow

retrieval grounded in verified org knowledge.

code is the source of truth. notion is treated as the human-readable layer that drifts. we keep them honest.
   ┌─────────────────┐         ┌─────────────────┐         ┌─────────────────┐
     github repos              notion pages              other sources  
     oauth · webhook           oauth · poll              v1.5 surface   
   └────────┬────────┘         └────────┬────────┘         └────────┬────────┘
                                                                  
            └─────────────┬─────────────┘                           │
                                                                   
                  [ celery ingest ]  ──── chunk · embed · upsert ───── 
                                                                   
                                                                   
            ┌─────────────────────────────────────┐                   │
                 postgres + pgvector               ◄── code is the truth
                 chunks · embeddings · meta      
            └─────────────────┬───────────────────┘
                              
                  [ retriever ]  ─── topK + rerank + intent shaping
                              
                              
                  [ cross-validator ]  flags notion ↔ code drift
                              
                              
                       claude sonnet    grounded generation
                              
                              
            ┌─────────────────┴───────────────────┐
               web uimcp pluginapi  
            └─────────────────────────────────────┘
// 03 ── specs

built like an instrument, not a wrapper.

every layer is replaceable. no hidden providers, no surprise dependencies. read the datasheet.

engram is a thin, opinionated stack assembled from boring components and one interesting idea — that cross-validation between code and docs catches staleness before it hurts.

the chat engine detects intent (explain vs. generate) and adjusts the prompt accordingly; the retrieval pipeline is shared. all llm calls route through a single provider abstraction so we can swap models without touching routes.

v0.1 · operationalsingle-tenantus-west-2 · eu-west-1tls 1.3
frontendnext typescript · tailwind · shadcn
backendfastapi python 3.12 · type-hints · ruff
databasepostgres 16 · pgvector
embedvoyage code fallback: text-embedding-3-small
llmclaude sonnet via anthropic sdk
queuecelery redis broker · async ingest
authgithub oauth jwt bearer (localstorage)
deployvercel + fly.io · sentry
mcpv1.5 30-day jwt · paste-install
// 04 ── plans

pay for indexed objects, not seats.

teams scale by repo size, not headcount. usage-aligned pricing reflects that. cancel anytime.
solo
$0/mo
1 repo · 1 workspace · best for evaluation
  • 50k indexed chunks
  • github + notion connect
  • community discord
  • self-host option
enterprise
talkto us
byo-cloud · sso · audit · custom retention
  • self-host or dedicated tenant
  • byo llm provider + key
  • sso · scim · audit log
  • private retrieval shaping
  • named support engineer

connect a repo. ask the real question.

start with one github repo and one notion workspace. engram indexes both, surfaces drift, and gives you cited answers in under six seconds from connect.