agent-qa (memidex)
A QA agent that remembers what tests proved about your app.
by vostride · Built this? Edit this page on GitHub
In short
agent-qa tests web and mobile apps from plain-language steps. Its memidex memory keeps product facts, cautions and working procedures with their evidence. Later runs receive relevant records, and new outcomes change how much they are trusted.
Who it’s for
Software teams and coding agents running repeatable tests of real app interfaces.
What it remembers
Product behavior, interface anchors, measurements, failure cautions and successful procedures.
- Memory lives
- Files in your project, with a local search index
- API keys
- A configured model provider or supported subscription authentication
- Works with
- CLI, MCP, Claude Code, Codex
- License
- Partly open sourceFSL-1.1-ALv2
- Price
- Local package; model-provider charges may apply
Under the hood
- Who writes memories
- After a test, fixed rules extract candidates and a model proposes changes that code checks.
- Where they're kept
- Markdown records and JSON evidence files; SQLite holds a rebuildable search index.
- How they're found
- Scoped briefs, full-text and vector search, anchor matches and bounded graph expansion.
- Extra AI calls
- Curation can call the configured model; optional replay and live canary checks add calls.
How it works
- 1TestRun a user journey and record evidence for each step.
- 2LearnTurn verified observations into candidate records.
- 3CheckValidate changes and promote eligible local candidates.
- 4RecallSupply relevant knowledge to the next test and update trust from its outcomes.
What stands out
- Evidence and learned facts travel together in reviewable record bundles.
- Code validates proposed lessons before they can influence later tests.
- A person can correct a verdict and the knowledge behind it through the CLI or MCP.
What it doesn’t do
- It is a testing harness, not a general personal-memory service.
- CI runs propose learned records but do not promote them; model access is required for full curation.
- Its current FSL-1.1-ALv2 license restricts competing use.
Measured against human memory
5 built, 4 partly, 2 not of the Field Guide’s 11 mechanisms
Encoding
Salience
Decides what's worth keeping, and how much, as a memory is written.
Fixed evidence rules select at most 50 candidates per run; curation can keep, confirm or skip them, without a separate importance score. source
Emotion
Gives memories written with feeling more weight.
Records describe test evidence, without emotional weighting. source
Storage
Forgetting
Lets unused memories fade over time on its own.
At serving time, trust is multiplied by 0.6 after 5 runs without confirmation, or 3 for measurements; the record remains on disk. source
Interference
When a new memory closely matches an old one, notices and lets the newer one win.
Curation looks for records on the same anchors and can supersede outdated facts, with checks against repeated conflicting changes. source
Retrieval
Retrieval
Remembering something makes that memory stronger.
Later observations can confirm a fact and increase its trust; simply serving or reading a record does not strengthen it. source
Association
Links memories, so recalling one brings its neighbours.
Search expands through neighboring graph records in one bounded hop, excluding high-degree hubs. source
Prospective
Brings a memory back at the right time without being asked.
Relevant knowledge returns in a test's starting brief and cautions can surface after failure; it is a test-context cue, not a reminder service. source
Transformation
Consolidation
Reworks memories between sessions: strengthening, merging, reorganising.
After the verdict, queued curation adds, confirms or replaces records for later runs, keeping an evidence trail; only a person retires an active record. source
Reconsolidation
Corrects a memory when it's contradicted, and keeps what it used to say.
Corrections supersede earlier records while preserving history; cleanup retains the three latest superseded generations and protects user records. source
Gist
Turns many specific moments into general knowledge.
Verified steps become reusable facts and successful flows become procedures; this is scoped test learning, not a separate cross-episode abstraction pass. source
Schemas
Forms beliefs about people and projects, and weighs new facts against them.
Keeps scoped facts and procedures rather than synthesizing a separate project-wide belief profile. source
Not every system is trying to be a human memory, so a hollow dot isn’t a flaw. It just says what the system doesn’t do. What the mechanisms are.
agent-qa (memidex)’s makers haven’t shared what they’re building next, so only today shows here. Makers can add their roadmap the same way they correct an entry.
Compare agent-qa (memidex) with the others, side by sideQuestions
- Who is agent-qa (memidex) for?
Software teams and coding agents running repeatable tests of real app interfaces.
- Where does agent-qa (memidex) keep its memories?
Files in your project, with a local search index. Markdown records and JSON evidence files; SQLite holds a rebuildable search index.
- Does agent-qa (memidex) forget?
Forgetting: built, basic. At serving time, trust is multiplied by 0.6 after 5 runs without confirmation, or 3 for measurements; the record remains on disk. source
- Does agent-qa (memidex) correct a memory when it turns out to be wrong?
Reconsolidation: built, basic. Corrections supersede earlier records while preserving history; cleanup retains the three latest superseded generations and protects user records. source
- Is agent-qa (memidex) open source?
Partly open source. License: FSL-1.1-ALv2. Code: github.com/vostride/agent-qa.
- Does agent-qa (memidex) work with Claude Code?
Yes. Works with: CLI, MCP, Claude Code, Codex.
- Does agent-qa (memidex) need an API key?
API keys: A configured model provider or supported subscription authentication.
- How much does agent-qa (memidex) cost?
Price: Local package; model-provider charges may apply.