Recall

Microsoft Recall is a feature released with Copilot+ PCs in 2024 that takes screenshots every few seconds and uses a local AI model to build a semantic index, allowing users to retrieve content that has appeared on their screen using natural language.

Microsoft's Recall feature, introduced in 2024, was forced to be delayed due to privacy controversy arising from its continuous screenshotting and AI-based indexing of user activity.

Security researchers discovered that Recall stored screenshots locally in an unencrypted SQLite database, sparking privacy controversy, after which Microsoft changed it to require users to actively enable it and added Windows Hello verification, encryption, and sensitive content filtering

When Microsoft's Recall feature was released in 2024, it faced strong backlash for continuously recording screenshots by default, and was forced to become an opt-in feature that is off by default, with its release delayed.

Recall is defined as the proportion of the returned Top-K results that truly belong to the set of true nearest neighbors.

IDP scoring should not report only a single metric; at minimum it should give Precision, Recall, and F1 together and state the counting rules for blanks, NULLs, checkboxes, and unreadable values.

After disconnecting FalkorDB, the Recall product degrades to an honest "I don't know" rather than silently falling back.

On a benchmark of 30 questions across five categories, graph retrieval scored 24.667 and flat RAG scored 19.251.

Developer Anshu built the open-source project Recall in three days.

Recall uses the graph database FalkorDB to carry the causal relationships of company knowledge, modeling events, PRs, commits, ADRs, services, people, teams, and chat messages as nodes connected by labeled edges, and supports Cypher queries.

Recall's core rule is: if the model cannot cite a certain node, it may not state the corresponding conclusion.

Recall builds a separate memory graph for each person, all running within the same FalkorDB instance.

Recall uses versioned memory: new facts mark old facts as superseded, and the old facts remain in the graph; the past is versioned but never rewritten.

When a user requests that a fact be completely forgotten, Recall performs the forgetting operation and issues an erasure receipt recording the identifier of the deleted fact, the deletion time, and the operator, but the receipt itself does not contain the fact's content.

Recall models approval actions as a node in the graph, persists approval state in FalkorDB, and has the executor read approved proposals from the graph database rather than relying on process memory, thereby achieving atomicity and recoverability of approvals.

The seven principles the author distilled include: every conclusion needs a citation; memory is scoped per person; the past is versioned rather than rewritten, and deletions must issue a receipt; the agent proposes and the human disposes; the graph is the foundation rather than a feature, and the product should honestly stop when it is disconnected; show the points lost, since a benchmark that hides its zeros is marketing while one that records its zeros is evidence; do not assert on LLM prose, assert on structured data.

The two hard zeros for graph retrieval in the benchmark were because the grader required exact wording while the model paraphrased; the README documents every point lost.

In a re-run validation, graph retrieval won again, scoring 25.0 versus 19.251.

In RAG scenarios, Recall is usually prioritized over Precision, because missing a piece of key evidence often directly leads to an incorrect answer, whereas retrieving a few extra irrelevant documents only increases context noise and token cost.

Microsoft Recall is a system-level feature in Windows 11 that periodically takes screenshots and builds a local index, allowing users to search past activity in natural language.