Free Coding Agent Memory: Start with the Hobby Plan

Free coding agent memory starts with the $0 Hobby plan: 10,000 adds, 1,000 retrieves, unlimited projects, and graph memory with no credit card. This guide shows how far that quota stretches a solo developer — and exactly when to move to Starter.

Hobby quotas: what the $0 plan includes

The Hobby plan costs $0 per month and requires no credit card. You get 10,000 add requests and 1,000 retrieve requests per month, unlimited projects, and graph memory. Graph memory is included on all plans from Hobby through Enterprise, so Hobby does not use a reduced recall model.

Billing is flat and monthly. There are no per-token fees and no overages: when you outgrow a quota, you move to the next tier. The service is fully managed, so you run no infrastructure. The API is a mem0-like API with five data-plane operations: POST /add, POST /search, GET /list, PUT /update, and DELETE /delete.

Starter holds 50,000 adds and 5,000 retrieves for $12, Growth holds 200,000 and 20,000 for $50, Pro holds 500,000 and 50,000 for $125, and Enterprise holds 1,000,000 and 100,000 for $260. Every tier keeps unlimited projects and graph memory, so upgrades change quota headroom only. The published comparison as of August 2026 describes the difference as flat billing instead of per-request overages, with matching quotas at $12, $50, and $125.

Data is isolated per organization with namespace-scoped vector search, and Audaxic never trains on customer data. Servers are located in the US and EU. Management create operations are limited to 100 per hour per IP, with RateLimit-Limit, RateLimit-Remaining, and RateLimit-Reset headers on responses. The pricing page lists current tiers, and the pricing documentation defines how each quota is counted.

How far free coding agent memory stretches for a solo developer

Size free coding agent memory by extraction, not API calls. An add request is counted per memory extracted from a message batch: one /add call that extracts five memories counts as five add requests. Your monthly adds equal memories extracted per session multiplied by sessions per day multiplied by working days.

Two sessions a day at 15 memories per session use about 30 memories a day. Over 22 working days that is 660 adds, under 7% of the Hobby quota. Three sessions a day at 25 memories each total 1,650 adds, under one-fifth of the 10,000 limit. These profiles fit Hobby with room to spare: daily sessions where your agent records preferences, project decisions, constraints, and gotchas. Five terse decisions cost five adds, while one long message that yields one decision costs one — length does not matter, extracted memories do.

Import months behave differently. The importer (npm install -g @audaxic/memory-import) backfills Claude Code, Codex, and OpenCode history with correct timestamps, and can extract hundreds of memories per session. One import week plus normal sessions can push a month past 10,000 adds. The boundary looks like this:

That entry point matches the case in why AI memory layers should be accessible to everyone: you start without a procurement step. Measure one normal week, multiply by four, and compare against 10,000. If you store durable facts only, Hobby covers a full solo month.

Unlimited projects: one memory per repo on Hobby

The word unlimited on Hobby applies to projects, and only to projects. Create one project per repository so each codebase keeps separate memory. That boundary protects recall quality: conventions for a Python API server do not belong in the same namespace as a React frontend.

Use one project and one ak_... project key per repo. Every agent working in that repo uses that key, so saves land in the right namespace and searches read from it. Data-plane calls authenticate with the project key against https://api.memory.audaxic.com. Workspace management uses scoped at_... auth tokens against the management plane instead, and management tokens never touch memory content.

Adding a fifth or fifteenth repo changes nothing on the bill. What changes is extraction volume: five quiet repos add almost nothing, while one large refactor can dominate the month's adds.

Keep keys out of the repo. Load the ak_... key from an environment variable, and use one key per repo so a search in one project never surfaces another project's memories. Page stored memories with GET /list, which supports limit and offset pagination. Prune stale facts with PUT /update and DELETE /delete/{memory_id}.

Set up Hobby memory in five minutes with a plugin or MCP

You have two paths. For Claude Code, Codex, or OpenCode, install the matching plugin. For other MCP-compatible clients, connect the remote MCP server.

Install the plugin for your agent:

# Claude Code
npx @audaxic/memory-claude install

# Codex
npx @audaxic/memory-codex install

# OpenCode
npm install -g @audaxic/memory-opencode

Then create a project, copy its ak_... key, and connect the plugin to that key. The quickstart walks through project creation, key handling, and the first store-and-recall cycle over REST. For Claude Code specifically, the Claude Code persistent memory setup guide shows the full five-minute flow with verification.

For MCP clients, the server URL is https://memory-mcp.audaxic.com/mcp over Streamable HTTP with a Bearer ak_... key. No local server or container is involved. The same URL covers Claude Desktop, Cursor, Windsurf, and any MCP-compatible client — Cursor connects via MCP only, since there is no Cursor plugin. Verify the binding after connecting: store one test fact about the current repo, open a fresh session, and search for it. If the new session recalls the fact, the project binding is correct. If it does not, the session is reading a different project than the one you wrote to — check the key.

Pin the X-Api-Version header to the current version (2026-09-25) when calling REST directly, so future version changes do not alter request shapes.

Where the 1,000 retrieves go: search counting

Retrieve requests use tiered counting: the first 10 memories returned count as 1 request, and each additional 10 counts as 1 more. A search returning 25 memories counts as 3 retrieve requests. Quota depends on result size, not just search count.

Three searches per session across three sessions a day is about nine searches daily. At 10 or fewer memories per search, that is 9 retrieve requests a day and roughly 198 across 22 working days — one-fifth of the Hobby quota. Broader searches returning 20 to 30 memories cost 2 to 3 requests each, so the same nine searches land near 400 to 600 for the month. Both profiles fit inside 1,000 with margin.

Session-start recall is the main retrieve consumer. The plugin injects relevant memories when a session opens, and each injection counts against quota. Mid-session searches add on top. Keep retrieves low with focused queries: ask for the convention or decision you need rather than pulling project history, and store atomic facts so small result sets answer the question.

Watch retrieves whenever extraction spikes. An import month raises adds far faster than retrieves, but broad searches on every prompt can invert that ratio. If GET /list shows thousands of stored memories and searches return full pages, narrow the query text and reduce the result limit before assuming the tier is too small.

When to move to Starter at $12

Starter costs $12 per month for 50,000 adds and 5,000 retrieves, with the same unlimited projects and included graph memory as Hobby. Billing stays flat with no per-request overages. The move is a quota decision, not a feature unlock: projects, graph memory, and endpoints do not change.

Three signals point to the switch. First, an import-heavy month: backfilling Claude Code, Codex, and OpenCode history can exceed 10,000 memories on its own, and $12 covers five times the Hobby add quota. Second, sustained extraction above roughly 500 memories per working day, which passes 10,000 across 22 days. Third, retrieve pressure above 1,000 requests from broad searches or many daily sessions, since Starter raises that ceiling to 5,000. As a reference point, mem0 lists 50,000 adds with 5,000 searches at $19 as of August 2026, so $12 holds matching quota for less. If none of those apply, free coding agent memory on Hobby stays the right default.

Beyond Starter the ladder keeps the same shape: Growth covers 200,000 adds and 20,000 retrieves at $50 per month, Pro covers 500,000 and 50,000 at $125, and Enterprise covers 1,000,000 and 100,000 at $260 with on-premise deployment available on request. Each step keeps unlimited projects, graph memory, and flat monthly billing. Compare all tiers in the published tier table before you decide.

Check the pricing documentation for the counting definitions, estimate from one measured week of your own sessions, and pick the smallest tier whose quotas clear your math with headroom.

FAQ

Written by Efe — Oct 11, 2026 · More articles · X @AudaxicTech