Hindsight is an open-source agent-memory architecture that turns conversation history into structured, queryable memory rather than relying only on semantically similar snippets. It separates memory into four logical networks and combines vector, keyword, graph, and temporal retrieval. That makes it a candidate for agents that need to track entities, relationships, and changing information—but its published benchmark scores are evaluation results, not guarantees for every agent or workflow.
What makes Hindsight a temporal memory graph?
A basic retrieval system can find past text that resembles a new query. An agent with longer-term memory may also need to answer who a statement was about, how facts relate, what happened before or after an event, and whether information has changed. Hindsight presents memory as a structured layer for that work: it incrementally organizes conversational material into a queryable memory bank, then uses a reflection layer to reason over it and update information in a traceable way.
In the Hindsight architecture, “temporal” and “graph” describe complementary concerns. Temporal handling is about information and relationships changing over time; graph-style organization preserves entities and their connections. Hindsight’s papers describe these as parts of its design, not as a universally accepted blueprint for agent memory. The exact schema and behavior depend on the project’s current implementation and configuration.
Four networks separate kinds of memory
The ACL 2026 demonstration paper and the authors’ 2025 preprint describe four logical networks:
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →#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.
- World: facts about the world.
- Experience: the agent’s experiences.
- Observation: synthesized summaries associated with entities.
- Opinion: evolving beliefs.
The separation is meant to help developers distinguish what an agent treats as a world fact from what it experienced, summarized, or believes. That distinction matters when evidence is incomplete or a person’s preferences have changed: a belief should not automatically be treated as an objective fact. These are Hindsight’s categories; the papers do not establish them as a standard every memory system follows.
How do retain, recall, and reflect work?
Hindsight names three operations for the memory lifecycle. The ACL abstract describes retrieval as combining multiple methods, with PostgreSQL and pgvector as the stated storage foundation.
| Operation | Role in the architecture | What it means for an agent |
|---|---|---|
| Retain | Ingests information into memory. | Conversation or other input becomes material the system can organize and later query. |
| Recall | Retrieves relevant memory. | The system looks for useful prior information rather than requiring the full interaction history in the prompt. |
| Reflect | Reasons over memory and can update it. | The agent can synthesize information and revise what it stores, with the paper describing updates as traceable. |
Why use several retrieval methods?
The ACL paper says Hindsight’s retrieval pipeline combines vector search, keyword matching, graph traversal, and temporal filtering. Each addresses a different kind of query: vector search can surface semantically related material; keyword matching can target explicit terms; graph traversal can follow entity relationships; and temporal filtering can narrow results by time. The papers describe the combined pipeline, but do not establish that every query uses every method in the same way.
For example, “What did the user say about moving?” may benefit from semantic matching, while “What changed after the move?” also depends on identifying the relevant entity and ordering information over time. This illustrates the design problem; it is not a claim about a specific Hindsight query result.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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.
How should an agent handle facts that change?
A temporal memory should not flatten every statement into a timeless fact. “The team is based in Boston” and “the team moved to Seattle in June” may both be true at different points. A useful system needs to preserve enough entity and time context to retrieve the right statement for a question such as “Where was the team based before June?”
Hindsight’s papers describe temporal, entity-aware memory and temporal filtering, alongside a reflection layer that can update information traceably. They do not, in the material cited here, specify a complete user-facing rule set for resolving every conflict, assigning effective dates, or choosing which statement to show when evidence disagrees. Do not assume a particular conflict-resolution policy without checking the current project documentation and testing the configuration you intend to deploy.
A practical design checklist
When evaluating Hindsight for a changing-information workflow, define the questions the agent must answer before choosing a memory design. Check whether the stored representation can preserve:
- The entity a statement refers to, including relationships to other entities.
- When a statement or relationship was valid, where that timing is available.
- The distinction between an observed statement, an inferred summary, and an evolving belief.
- Enough evidence or provenance to inspect why the agent returned or changed information.
These are evaluation criteria for a temporal memory system, not a claim that every item is exposed through a particular Hindsight API or schema. The project’s README points to current documentation; implementation details should be verified there.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →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.
How does Hindsight compare with vector retrieval and temporal graphs?
A vector database is a storage and retrieval component, not by itself a complete agent-memory policy. A temporal knowledge graph emphasizes entities, relationships, and change over time. Hindsight presents a broader memory architecture that combines multiple retrieval methods with separate memory networks and retain, recall, and reflect operations. These categories overlap in practice, so compare actual implementations and workloads rather than labels alone.
| Comparison axis | Hindsight | Vector retrieval alone | Zep Graphiti |
|---|---|---|---|
| Representation | Four logical networks for world facts, experiences, synthesized entity summaries, and evolving beliefs (Hindsight authors, 2025 preprint; ACL, 2026). | Not specified by the Hindsight papers; vector retrieval alone does not define an agent’s fact-versus-belief policy. | Temporally aware knowledge graph that combines conversational information with structured business data (Zep authors, 2025 preprint). |
| Time and relationships | Paper describes temporal filtering and an entity-aware memory layer; exact update policy depends on implementation. | Not established by vector search alone. | Paper describes retaining historical relationships. |
| Retrieval methods | Vector search, keyword matching, graph traversal, and temporal filtering (ACL, 2026). | Vector search; other methods depend on the system built around it. | Not stated here as a directly comparable retrieval-method list. |
| Storage and distribution | PostgreSQL with pgvector is named in the ACL abstract. The publication reports an MIT-licensed open-source project, Python package, and Docker image. | Varies by database and application; no single deployment requirement applies. | Not stated here in a directly comparable deployment format. |
| Latency, cost, and setup burden | Not stated as comparable values in the cited Hindsight papers. | Not stated; depends on model, database, and system design. | Not stated here as directly comparable values. |
The comparison is about architectural emphasis, not a claim that one system always outperforms another. Hindsight’s paper discusses MemGPT, Zep, and Mem0 in the context of its feature set; that does not establish that Hindsight is the only system with any given capability. Zep’s Graphiti preprint describes a related temporal-graph approach, but the sources summarized here do not provide aligned deployment, latency, and cost measurements across these systems.
What do Hindsight’s benchmark scores show?
The numbers below are reported by the Hindsight authors or the ACL publication. They are not independent cross-vendor audit results, and they should not be treated as a universal ranking. The model configuration and benchmark matter:
| Source and configuration | Reported result | Scope |
|---|---|---|
| Hindsight authors’ 2025 preprint, open-source 20B model | 83.6% on LongMemEval | The same preprint reports 39% for its full-context baseline using the same backbone. |
| Hindsight authors’ 2025 preprint, larger-backbone configuration | 91.4% on LongMemEval | The preprint’s reported result for that configuration. |
| Hindsight authors’ 2025 preprint, stronger configuration | 89.61% on LoCoMo | The preprint also reports 75.78% for the strongest prior open system in its comparison. |
| Association for Computational Linguistics, 2026 publication, 20B open-source model | 83.6% on LongMemEval; 83.2% on LoCoMo | Reported results for the stated model configuration. |
| Association for Computational Linguistics, 2026 publication, Gemini-3 Pro | 91.4% on LongMemEval | Reported result for that model configuration. |
| Zep authors’ 2025 Graphiti preprint | 94.8% versus 93.4% on DMR | Zep authors’ result in their evaluation context. It is not directly comparable to Hindsight’s percentages without aligned models, prompts, data, and scoring. |
Different backbones can produce different scores, even within one system. The comparison of Hindsight’s 20B result with its full-context baseline is specifically the authors’ same-backbone LongMemEval setup; it is not evidence that memory will beat full-context prompting in every application.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsRank #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
Why scores may not predict agent performance
In March 2026, the Hindsight team argued that LongMemEval and LoCoMo remain useful but may not distinguish memory architectures well when large-context models can fit the evaluation material. The team also said those datasets emphasize chatbot-style conversational recall more than multi-step agent tasks. That is the project’s assessment of benchmark limits, not an independent consensus finding.
Before comparing two systems, align the evaluation conditions and record:
- The exact model and prompt used.
- What the baseline includes.
- The benchmark split and scoring procedure.
- Latency and inference cost as well as accuracy.
- Setup and tuning effort.
- Whether the test resembles the agent’s real workflow, particularly if it must carry out multi-step tasks.
The Hindsight team’s benchmark commentary stresses publishing methodology because judge prompts, answer-generation prompts, and model choice can materially change measured accuracy. A score without those details is a weak basis for a deployment decision.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Can you run Hindsight locally?
The ACL 2026 publication says Hindsight is open source under the MIT license and distributed as a Python package and Docker image. It gives the Python package command as pip install hindsight-all. The publication also reports production use at Fortune 500 enterprises; that is an author-reported statement, not a disclosure of customer names or deployment details.
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.
The package command alone is not a complete setup guide. Requirements, supported models, configuration, and current deployment steps can change, so consult the project’s current documentation before installing or designing an integration. The project positions Hindsight for conversational and autonomous task-oriented agents, especially where feedback should change future behavior; that is the project’s intended-use description, not independent proof of outcomes.
When is Hindsight worth evaluating?
Hindsight is most relevant when an agent needs more than recall of similar passages: the application must distinguish facts from beliefs or experience, follow relationships among entities, or retrieve information with time in mind. Its architecture combines those concerns with a reflection operation, while its published scores provide evidence on specific conversational-memory benchmarks—not a guarantee for every production task.
For a real selection, test representative queries and updates from the intended workflow, inspect whether returned information is traceable, and measure accuracy, latency, inference cost, and operational effort under the same conditions for each candidate. A benchmark result is useful context; the fit to the agent’s actual work is the decision.
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.
