Typed events are a better canonical record for an agent that must answer what changed, when it changed, who or what supplied the information, and which scope it belongs to. Embeddings remain useful as a derived index for fuzzy semantic recall. The defensible design is not an either-or choice: keep versioned, auditable events as the source of truth, then build lexical, vector, graph, or other projections for the questions users actually ask.
What typed-event memory stores
A typed event is a durable record of something that happened or a fact that changed. Instead of reducing an interaction to an opaque numerical vector, the record names its meaning and carries the metadata needed to interpret it later. A practical event commonly includes:
As an Amazon Associate I earn from qualifying purchases.
- Stable identifier: an ID that remains addressable when projections or indexes are rebuilt.
- Event type: a controlled name such as
PreferenceRecorded,TaskCompleted,PlanRejected, orContactUpdated. - Timestamp: event time, and often ingestion time as a separate field when messages arrive late.
- Scope: the user, agent, workspace, tenant, project, or conversation allowed to retrieve it.
- Subject and relationships: IDs for the person, task, document, or other entities affected.
- Provenance: the source message, tool result, human confirmation, or upstream system that supports the record.
- Payload: the typed values that describe the change.
- Schema version and lifecycle state: the format of the event, retention status, and any superseding or correcting event.
For example, an event could state that a user changed a delivery address on 2026-10-18, identify the account and address version, record whether the change was confirmed by the user or inferred from a message, and point to the earlier address that it supersedes. A later read model can show the current address, while the event history preserves how that state was reached.
Recommended Free Tools
“Append-only” does not mean “never correct.” It means the original observation is retained and a correction, retraction, or deletion request is represented explicitly. That distinction lets an application answer both “What is true now?” and “What did the system believe before the correction?” without silently rewriting history.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Why embeddings alone are a poor authority for agent memory
An embedding converts text or another object into a vector so that a query can retrieve items that are semantically close even when the wording differs. That is valuable for open-ended recall, but similarity ranking is not a complete memory model.
Exact time and order
A question such as “What changed after the contract was renewed but before the invoice was paid?” requires ordered timestamps and deterministic boundaries. A vector search may return relevant passages, but similarity by itself does not guarantee a complete or correctly ordered timeline. First-class event time, sequence information, and explicit time filters do.
Scope and authorization
Agents often operate across users, tenants, projects, and tools. A typed record can be filtered by scope before any language model sees it. A nearest-neighbor result is not an authorization policy; the index must carry enforceable scope metadata and the query path must apply it consistently.
Audit and provenance
When an agent gives consequential advice, operators need to identify the source of the remembered fact, the version of the data, and the transformation that produced a summary. A vector distance says which items were close in an embedding space, not why a fact was accepted, whether it was later disputed, or which source supports it.
Corrections and current state
Embeddings do not naturally express that one belief supersedes another. Re-embedding a rewritten summary can hide the prior state and make it difficult to explain a previous answer. Explicit event types such as PreferenceRevoked or TaskCancelled let projections apply deterministic rules for current state while preserving the change history.
Relationships and multi-hop questions
“Which open projects depend on the component that failed in last week’s deployment?” is a relationship query. Typed entity IDs, relational joins, or a graph projection represent those edges directly. Nearby vectors may retrieve descriptions of the projects and component, but they do not guarantee that the dependency path is complete.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Memory is a set of subtypes, not one homogeneous database
Microsoft’s multi-agent architecture guidance makes the central point plainly: “Storage strategy should be chosen per memory sub-type, not for memory as a whole.” A durable user profile, an episodic timeline, and a searchable document corpus have different access patterns and retention rules.
Free tools Windows power users keep installed
One-click scans. No signup required.
| Memory subtype | Typical question | Suitable authority | Useful derived views |
|---|---|---|---|
| Profile or preferences | What communication setting is active for this user? | Structured record with explicit update and validity rules | Small text or vector index for finding related profile notes |
| Episodic timeline | What happened, in what order, during this case? | Typed events with timestamps, scope, and provenance | Time-range projection, summaries, lexical search, optional vectors |
| Knowledge or document corpus | Which passages discuss this concept? | Document store and source metadata | Lexical and vector retrieval, citation-oriented chunks |
| Relationship data | How are these people, services, and tasks connected? | Relational or graph structures | Text and vector indexes over node and edge descriptions |
A hybrid arrangement is common: structured persistence for the profile or event history plus a vector index for episodic or document recall. It also creates consistency work. Every derived index needs an update path, a failure policy, and a rebuild procedure; otherwise the system accumulates stale or contradictory views.
A canonical event log with rebuildable retrieval views
The most robust interpretation of “typed events instead of embeddings” is to make events authoritative and treat search indexes as projections. A typical flow has four layers:
- Canonical history: validated events are stored with IDs, timestamps, scope, provenance, schema version, and retention metadata.
- Read projections: consumers fold events into current-profile tables, timelines, counters, or materialized views that answer exact queries quickly.
- Lexical and vector indexes: selected events, summaries, or source text are indexed for keyword and fuzzy semantic retrieval.
- Context assembly: a query planner combines deterministic filters, relationship traversal, and semantic candidates, then returns records with provenance to the agent.
An index is genuinely rebuildable only when the original events, the transformation code, the embedding-model identity, chunking rules, and configuration needed to reproduce it are retained. The IETF document titled Architecture and Data Model for Persistent Memory in Agentic Systems describes this separation between event history and derived embedding spaces as a proposal; it is an individual Internet-Draft, not an adopted standard.
Rebuildability changes operational decisions. If a vector service loses data, operators can regenerate it from canonical records. If an embedding model changes, the old and new spaces can be identified and migrated deliberately instead of being mixed under one index name. During a rebuild, the application can fall back to exact projections or lexical search rather than treating missing vectors as missing memory.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →How a typed-memory pipeline works
1. Decide what deserves durable memory
Do not append every token or transient chat turn automatically. Define extraction rules for durable preferences, commitments, outcomes, corrections, and other events that matter after the conversation ends. Keep ephemeral context in a short-lived store when long-term retention has no user or product value.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
2. Extract and validate an event
An extraction model can propose an event, but validation should check required fields, allowed event types, scope, timestamps, referenced entity IDs, and confidence or confirmation requirements. A human confirmation or trusted tool result may be required before a high-impact fact becomes durable.
3. Append the canonical record
Write the event once with an idempotency key so retries do not duplicate it. Record both the event timestamp and ingestion timestamp when ordering can be ambiguous. Store provenance that an operator can follow back to the source.
4. Update projections
Consumers apply events to a current-state view, a time-ordered timeline, or a relationship index. Make processing idempotent and observable: record the last event position, failed handlers, and projection version. A projection should be disposable and recreatable from the log.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
5. Add semantic indexing selectively
Embed only records for which wording variation creates a real retrieval problem. Index a concise event rendering or a trusted summary, not arbitrary hidden chain-of-thought or sensitive fields. Store the event ID, scope, timestamp, and embedding-space identity beside the vector.
6. Plan queries before retrieval
Classify the request. Exact time, type, scope, status, and relationship constraints should become database or graph filters first. Use lexical or vector retrieval for the remaining fuzzy portion. Merge candidates, remove superseded or unauthorized records, and preserve source references in the context sent to the model.
7. Handle corrections, deletion, and retention
A correction event can supersede a prior belief in projections while retaining an auditable link. Privacy requirements may require deletion or redaction of both canonical data and every derived index. Define retention periods and tenant boundaries independently of the search technology; a vector index must not become a loophole around deletion policy.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
Typed events and embeddings compared by engineering concern
| Concern | Typed events or structured records | Embedding index | Design implication |
|---|---|---|---|
| Representation | Named facts or changes with fields and relationships | Numerical representation ranked by similarity | Declare which layer is canonical and which is derived. |
| Exact temporal questions | Natural fit when timestamps and event types are first-class | Similarity does not ensure complete ordering or exact boundaries | Preserve timestamps and apply deterministic time filters. |
| Fuzzy wording | Needs lexical search, query translation, or a model | Strong fit for semantic recall across varied phrasing | Add vectors only where fuzzy recall improves a measured task. |
| Audit and provenance | Source, scope, revisions, and schema can be explicit | Similarity ranking is not a provenance model | Return source references with every retrieved memory. |
| Relationship traversal | Typed links, joins, or graph edges support multi-hop queries | Nearby vectors are not explicit relational edges | Use relational or graph projections for dependency questions. |
| Evolution | Requires schema migration, upcasting, and projection rebuilds | Model or vector-space changes can require re-indexing | Version both event schemas and embedding spaces. |
| Operations | Schema, consumers, retention, and projections must be maintained | Embedding generation, vector infrastructure, and freshness add burden | Choose the simplest combination that answers the workload. |
The lifecycle costs that event-sourced memory introduces
A 2021 study of 19 event-sourced systems identified recurring practitioner challenges: system evolution, a steep learning curve, limited available technology, projection rebuilding, and data privacy. Those findings are warnings to design for, not proof that every agent-memory project will incur identical costs.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteSchema evolution
Event names and payloads live longer than the code that first emitted them. Use explicit schema versions, compatibility rules, and migration tests. Common tactics include upcasting old events when they are read, maintaining weak schemas while formats stabilize, transforming data in place where policy permits, or copying and transforming into a new stream. Pick one policy per event family and document when old versions stop being accepted.
Projection rebuilds
A new query, bug fix, or authorization rule may require replaying the history. Keep projection code deterministic, record projection versions, and test replay on representative histories. For large streams, support checkpoints and parallel rebuilds, but do not sacrifice the ability to verify results against a known event range.
Privacy and deletion
Immutable history can conflict with data-minimization obligations. Classify fields, encrypt sensitive payloads, separate identifying data from analytical events where possible, and define cryptographic erasure or redaction procedures before launch. Verify that summaries, lexical indexes, vector stores, caches, and backups follow the same policy.
Operational expertise
Teams must understand ordering, idempotency, consumer lag, failed events, and replay safety. A simpler structured table may be preferable when the workload does not need historical reconstruction. Event sourcing is a means to obtain explicit history, not a requirement for every memory feature.
When to add vector search
Use vectors when the user’s wording is likely to differ substantially from the stored wording and a relevant result cannot be found reliably with fields or keywords. Examples include recalling a past discussion about “the blue sedan issue” when the event text says “vehicle color mismatch,” or locating related policy passages across varied terminology.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Do not use vector similarity as a substitute for authorization, status, chronology, or correction logic. Apply scope and time filters before ranking where the vector system supports them, or retrieve a bounded candidate set and enforce those constraints in a trusted service. Keep the event ID in each vector record so the final response can hydrate the authoritative object and show provenance.
Text search can be a cheaper and more transparent first step for names, identifiers, error codes, and exact phrases. A graph or relational query is better for explicit dependencies. The query planner can route each clause to the mechanism that represents it directly, then combine the results under a fixed context-token budget.
How to evaluate the architecture without a misleading benchmark
There is no defensible universal percentage for typed events versus embeddings. The relevant result depends on workload, extraction quality, index freshness, model choice, and the cost of false memories. A useful evaluation set should include:
- Exact temporal questions with known complete answer sets.
- Scope and tenant-isolation tests, including adversarial cross-scope queries.
- Fuzzy paraphrases that measure semantic recall separately from exact lookup.
- Correction and deletion cases that verify superseded data never wins.
- Relationship and multi-hop questions with a known graph of entities.
- Replay tests measuring projection correctness after schema changes.
- Operational measurements for freshness, rebuild duration, storage, model calls, and context size.
- Human review of provenance and the agent’s ability to explain why a memory was used.
Report results for the specific workload and date, and identify the embedding model, event schema, corpus size, filters, and hardware. The fact that an event-sourcing study examined 19 systems is a description of its sample, not a performance claim about agent memory.
A practical decision checklist
- Do users ask “when,” “before,” “after,” “what changed,” or “who confirmed it”? If so, make timestamps, event types, and provenance first-class.
- Must the system explain or audit an answer? Keep an authoritative structured record and source references.
- Are queries mostly paraphrases over unstructured documents? Add lexical or vector retrieval.
- Do questions traverse explicit dependencies? Add relational or graph projections.
- Will schemas, embedding models, or retention rules change? Version them and test migrations before production.
- Can every derived index be rebuilt from retained data and deterministic transformations?
- Can corrections, user deletion, tenant isolation, and access revocation reach every projection?
- Is the added infrastructure justified by a measured retrieval problem rather than by the popularity of vector databases?
Bottom line for an agent-memory design
Typed events are the right authority when memory must be inspectable, time-aware, scoped, correctable, and replayable. Embeddings solve a different problem: finding semantically related material despite wording differences. Store the durable history in an explicit, versioned form; derive exact read models and add vector or lexical indexes where tests show they improve recall. This division preserves auditability without giving up flexible search, while making the costs of evolution, privacy, and operations visible before they become production failures.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
