Account context & memory

Teach Mavera your business once — every reply, plan, and study uses it. Editable, queryable, and self-updating.

Mavera keeps a structured memory of your account: your business, product, customers, competitors, style, and preferences. It's compiled into Mave's system prompt on every turn, so replies and study plans are specific to your world without re-explaining it each time.

How it's structured

Each entry is one declarative fact (max 500 chars), in one of three scopes and seven categories:


Scopes	`account` (whole workspace, 6,000 chars) → `project` (3,000) → `conversation` (1,500). Compilation stacks them in that order.
Categories	`business`, `product`, `customers`, `competitors`, `style`, `preferences`, `facts`
Sources	`user` (you wrote it) or `learned` (Mave extracted it from a conversation)

Budgets are enforced on the final state of a batch — a single call can remove stale entries and add new ones even when an add alone would overflow.

Read & search

GET /api/v1/context

curl "$BASE/api/v1/context?scope=account&q=competitors" \
  -H "x-api-key: $MAVERA_API_KEY"

Returns entries plus live per-scope budget usage. q= is a keyword filter over category + content.

Edit (batch, atomic)

POST /api/v1/context

{
  "scope": "account",
  "ops": [
    { "action": "add", "category": "product", "content": "Waggle: smart dog collar, $99 + $8/mo subscription" },
    { "action": "replace", "oldText": "mid-market CFOs", "content": "Primary buyers are heads of RevOps at 50-500 seat SaaS companies", "category": "customers" },
    { "action": "remove", "oldText": "discontinued the free tier" }
  ]
}

replace/remove target by id or a unique oldText substring. Also available: PUT (edit one entry), DELETE ?id= (archive — soft delete, revertible in the DB).

See exactly what the model sees

GET /api/v1/context/compile

curl "$BASE/api/v1/context/compile?projectId=proj_123" \
  -H "x-api-key: $MAVERA_API_KEY"

Returns the byte-for-byte context block injected into Mave's system prompt for that scope combination. No other research tool exposes this.

Self-updating memory

After conversations, a low-effort extractor reviews the exchange against existing entries and applies small batches of ops (max 5 per turn) under strict rules:

Declarative facts only — "User prefers terse answers", never instructions
Corrections replace the stale entry in the same batch — they don't pile up
Never stored: task progress, one-off study parameters, run results, anything stale within a week
The empty op list is the normal outcome — the bar is "would a colleague joining next month need this?"

Learned entries surface in the chat as a "Noted for next time" chip and carry a LEARNED badge in the Context page — edit or delete them like anything else. Pin an entry (📌) and the learner can never modify or remove it; only you can.

In the app

Context (sidebar → Advanced): grouped by category, inline edit, search, pin, budget meter, and the "What Mave sees" toggle showing the compiled block.