mirror of
https://github.com/nesquena/hermes-webui.git
synced 2026-05-27 12:10:40 +00:00
fix(memory): address 8 remaining feedback items
1. OpenViking link: OpenViking/OpenViking (404) → volcengine/OpenViking (22.4K stars, verified) 2. OpenViking L0/L1/L2 description: rewritten as resolution depth (50/500/full tokens), not temperature hierarchy — this is the actual mechanism behind the token savings 3. Hindsight benchmark: standardized to 91.4–94.6% everywhere (was mixed 91–94% / 91–94.6%) 4. Hindsight link: provider name now points to hindsight.vectorize.io (product page); GitHub link kept in meta line for developers 5. ByteRover CLI license: noted as custom license (campfirein/byterover-cli uses 'Other'), not MIT as was implied 6. Mobile column visibility: Paid from now stays visible on mobile; Tools column hidden instead (Stars already hidden) — pricing is more decision-relevant than tool count 7. OG/Twitter meta tags added: og:title, og:description, og:image, twitter:card, twitter:title, twitter:description, twitter:image — shares will preview correctly 8. Holographic fact_store: card now notes fact_store exposes 9 actions (add, search, probe, related, reason, contradict, update, remove, list)
This commit is contained in:
Hermes Agent
+32
-23
@@ -13,6 +13,15 @@
|
||||
<meta charset="UTF-8">
|
||||
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
||||
<title>Memory Providers for AI Agents — 2026 Guide</title>
|
||||
<meta property="og:title" content="Memory Providers for AI Agents — 2026 Guide">
|
||||
<meta property="og:description" content="Compare 8 agent memory providers — Mem0, Hindsight, ByteRover, Supermemory, Holographic, OpenViking, Honcho, RetainDB. Free tiers, benchmarks, hosting, and use-case picks.">
|
||||
<meta property="og:type" content="website">
|
||||
<meta property="og:url" content="https://get-hermes.ai/memory/">
|
||||
<meta property="og:image" content="https://get-hermes.ai/images/ui-hero.png">
|
||||
<meta name="twitter:card" content="summary_large_image">
|
||||
<meta name="twitter:title" content="Memory Providers for AI Agents — 2026 Guide">
|
||||
<meta name="twitter:description" content="Compare 8 agent memory providers — free tiers, benchmarks, architecture, and a straight pick for every use case.">
|
||||
<meta name="twitter:image" content="https://get-hermes.ai/images/ui-hero.png">
|
||||
<meta name="description" content="Which memory provider should you use with your AI agent? Compare Mem0, Hindsight, ByteRover, Supermemory, and more — free tiers, hosting options, benchmarks, and use-case picks.">
|
||||
<style>
|
||||
:root {
|
||||
@@ -174,7 +183,7 @@
|
||||
.t-bench small { color: var(--text-muted); display: block; font-size: 0.7rem; }
|
||||
/* hide less-critical columns on small screens */
|
||||
@media (max-width: 820px) { .col-arch { display: none; } }
|
||||
@media (max-width: 640px) { .col-stars { display: none; } .col-paid { display: none; } }
|
||||
@media (max-width: 640px) { .col-stars { display: none; } .col-tools { display: none; } }
|
||||
|
||||
/* BENCHMARK CALLOUT */
|
||||
.bench-callout {
|
||||
@@ -366,7 +375,7 @@
|
||||
<th>Hosting</th>
|
||||
<th>License</th>
|
||||
<th class="col-stars">Stars</th>
|
||||
<th>Tools</th>
|
||||
<th class="col-tools">Tools</th>
|
||||
<th>Benchmark</th>
|
||||
</tr>
|
||||
</thead>
|
||||
@@ -379,19 +388,19 @@
|
||||
<td>Cloud + self-host</td>
|
||||
<td><span class="pill pill-blue">Apache 2.0</span></td>
|
||||
<td class="t-stars col-stars">51.4K ⭐</td>
|
||||
<td>3</td>
|
||||
<td class="col-tools">3</td>
|
||||
<td class="t-bench">LongMemEval-S <strong>67.6%</strong></td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td class="t-name"><a href="https://github.com/vectorize-io/hindsight" target="_blank" rel="noopener">Hindsight</a></td>
|
||||
<td class="t-name"><a href="https://hindsight.vectorize.io" target="_blank" rel="noopener">Hindsight</a></td>
|
||||
<td>Best benchmarks, coding</td>
|
||||
<td><span class="pill pill-green">Full local, free</span></td>
|
||||
<td class="col-paid">$15/M retain<br><small style="color:var(--text-muted)">$0.75/M recall · $3/M reflect</small></td>
|
||||
<td>Local + cloud</td>
|
||||
<td><span class="pill pill-blue">MIT</span></td>
|
||||
<td class="t-stars col-stars">2.4K ⭐</td>
|
||||
<td>4</td>
|
||||
<td class="t-bench">LongMemEval <strong>91–94%</strong><small>BEAM 64.1% · LoCoMo 89.6%</small></td>
|
||||
<td class="col-tools">4</td>
|
||||
<td class="t-bench">LongMemEval <strong>91.4–94.6%</strong><small>BEAM 64.1% · LoCoMo 89.6%</small></td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td class="t-name"><a href="https://byterover.dev" target="_blank" rel="noopener">ByteRover</a></td>
|
||||
@@ -401,7 +410,7 @@
|
||||
<td>Local + cloud</td>
|
||||
<td><span class="pill pill-amber">Partial OSS</span></td>
|
||||
<td class="t-stars col-stars">4.2K ⭐</td>
|
||||
<td>3</td>
|
||||
<td class="col-tools">3</td>
|
||||
<td class="t-bench">LoCoMo <strong>92.2%</strong><small>single-hop 95.4% · temporal 94.4%</small></td>
|
||||
</tr>
|
||||
<tr>
|
||||
@@ -412,7 +421,7 @@
|
||||
<td>Cloud only</td>
|
||||
<td><span class="pill pill-red">Proprietary</span></td>
|
||||
<td class="t-stars col-stars">~18K ⭐</td>
|
||||
<td>4</td>
|
||||
<td class="col-tools">4</td>
|
||||
<td class="t-bench">LongMemEval <strong>81.6%</strong><small>with GPT-4o</small></td>
|
||||
</tr>
|
||||
<tr>
|
||||
@@ -423,18 +432,18 @@
|
||||
<td>Local only</td>
|
||||
<td><span class="pill pill-blue">MIT</span></td>
|
||||
<td class="t-stars col-stars">—</td>
|
||||
<td>2</td>
|
||||
<td class="col-tools">2</td>
|
||||
<td class="t-bench">—</td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td class="t-name"><a href="https://github.com/OpenViking/OpenViking" target="_blank" rel="noopener">OpenViking</a></td>
|
||||
<td class="t-name"><a href="https://github.com/volcengine/OpenViking" target="_blank" rel="noopener">OpenViking</a></td>
|
||||
<td>On-prem / air-gapped</td>
|
||||
<td><span class="pill pill-green">Self-host, free</span></td>
|
||||
<td class="col-paid">—</td>
|
||||
<td>Self-host only</td>
|
||||
<td><span class="pill pill-amber">AGPL-3.0</span></td>
|
||||
<td class="t-stars col-stars">~17.9K ⭐</td>
|
||||
<td>5</td>
|
||||
<td class="col-tools">5</td>
|
||||
<td class="t-bench">—</td>
|
||||
</tr>
|
||||
<tr>
|
||||
@@ -445,7 +454,7 @@
|
||||
<td>Cloud + self-host</td>
|
||||
<td><span class="pill pill-amber">AGPL-3.0</span></td>
|
||||
<td class="t-stars col-stars">414 ⭐</td>
|
||||
<td>4</td>
|
||||
<td class="col-tools">4</td>
|
||||
<td class="t-bench">—</td>
|
||||
</tr>
|
||||
<tr>
|
||||
@@ -456,7 +465,7 @@
|
||||
<td>Cloud only</td>
|
||||
<td><span class="pill pill-red">Proprietary</span></td>
|
||||
<td class="t-stars col-stars">—</td>
|
||||
<td>5</td>
|
||||
<td class="col-tools">5</td>
|
||||
<td class="t-bench">—</td>
|
||||
</tr>
|
||||
</tbody>
|
||||
@@ -504,8 +513,8 @@
|
||||
<div class="pc-header">
|
||||
<div class="pc-icon">🔬</div>
|
||||
<div>
|
||||
<div class="pc-name"><a href="https://github.com/vectorize-io/hindsight" target="_blank" rel="noopener">Hindsight</a></div>
|
||||
<div class="pc-meta">MIT · vectorize-io/hindsight · 2.4K ⭐</div>
|
||||
<div class="pc-name"><a href="https://hindsight.vectorize.io" target="_blank" rel="noopener">Hindsight</a></div>
|
||||
<div class="pc-meta">MIT · <a href="https://github.com/vectorize-io/hindsight" target="_blank" rel="noopener" style="color:inherit">vectorize-io/hindsight</a> · 2.4K ⭐</div>
|
||||
</div>
|
||||
</div>
|
||||
<p class="pc-desc">TEMPR architecture: four parallel retrieval strategies — temporal, entity, metadata, and BM25 for exact keyword matches. Strong at structured technical recall: port numbers, error codes, service names, deployment configs. Three-stage pipeline: retain (ingest) → recall (retrieve) → reflect (synthesize across stored knowledge).</p>
|
||||
@@ -516,7 +525,7 @@
|
||||
</div>
|
||||
<div class="pc-rows">
|
||||
<div class="pc-row"><span class="pc-row-label">Pricing</span><span class="pc-row-val">Full local free · Cloud: $15/M retain · $0.75/M recall · $3/M reflect</span></div>
|
||||
<div class="pc-row"><span class="pc-row-label">Benchmarks</span><span class="pc-row-val">LongMemEval <strong>91–94.6%</strong> · BEAM 64.1% · LoCoMo 89.6%</span></div>
|
||||
<div class="pc-row"><span class="pc-row-label">Benchmarks</span><span class="pc-row-val">LongMemEval <strong>91.4–94.6%</strong> · BEAM 64.1% · LoCoMo 89.6%</span></div>
|
||||
<div class="pc-row"><span class="pc-row-label">Config key</span><span class="pc-row-val"><a href="https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers">memory_provider: hindsight</a></span></div>
|
||||
</div>
|
||||
<div class="pc-best"><strong>Best for:</strong> coding agents, privacy-first local setups, anyone who wants the best published benchmark scores. Highest LongMemEval score of any provider listed.</div>
|
||||
@@ -527,7 +536,7 @@
|
||||
<div class="pc-icon">🤖</div>
|
||||
<div>
|
||||
<div class="pc-name"><a href="https://byterover.dev" target="_blank" rel="noopener">ByteRover</a></div>
|
||||
<div class="pc-meta">CLI: open source · cloud: proprietary · 4.2K ⭐</div>
|
||||
<div class="pc-meta">CLI: open source (custom license) · cloud: proprietary · 4.2K ⭐</div>
|
||||
</div>
|
||||
</div>
|
||||
<p class="pc-desc">Leads the LoCoMo benchmark — specifically designed for multi-hop and temporal reasoning across long conversation histories. Local CLI is open source and free. Cloud sync for cross-device persistence costs $19/mo. Built with coding agents as the primary use case.</p>
|
||||
@@ -574,7 +583,7 @@
|
||||
<div class="pc-meta">MIT · built into Hermes · local SQLite</div>
|
||||
</div>
|
||||
</div>
|
||||
<p class="pc-desc">Uses Holographic Reduced Representations (HRR) algebra on a local SQLite + FTS5 store. Zero external dependencies — no API keys, no network calls, no Docker. Memory lives in a single file in your Hermes home directory. The most private option by definition.</p>
|
||||
<p class="pc-desc">Uses Holographic Reduced Representations (HRR) algebra on a local SQLite + FTS5 store. Zero external dependencies — no API keys, no network calls, no Docker. Memory lives in a single file in your Hermes home directory. The most private option by definition. The <code style="font-family:monospace;font-size:0.82em">fact_store</code> tool exposes 9 actions: add, search, probe, related, reason, contradict, update, remove, list.</p>
|
||||
<div class="pc-tags">
|
||||
<span class="pill pill-green">Fully local, free</span>
|
||||
<span class="pill pill-blue">Local only</span>
|
||||
@@ -592,11 +601,11 @@
|
||||
<div class="pc-header">
|
||||
<div class="pc-icon">🏛️</div>
|
||||
<div>
|
||||
<div class="pc-name"><a href="https://github.com/OpenViking/OpenViking" target="_blank" rel="noopener">OpenViking</a></div>
|
||||
<div class="pc-name"><a href="https://github.com/volcengine/OpenViking" target="_blank" rel="noopener">OpenViking</a></div>
|
||||
<div class="pc-meta">AGPL-3.0 · self-hosted · ~17.9K ⭐</div>
|
||||
</div>
|
||||
</div>
|
||||
<p class="pc-desc">Tiered loading system: L0 (hot, in-memory), L1 (warm, local cache), L2 (cold, storage). Cuts token usage 80–90% by loading only what's relevant at each level. Self-hosted only, AGPL license. Requires Docker and an LLM provider for extraction. No cloud service.</p>
|
||||
<p class="pc-desc">Tiered context loading by resolution depth: L0 loads ~50-token abstracts, L1 loads ~500-token overviews, L2 loads full content on demand. Only the detail level needed for each query gets pushed into the context window — that's the mechanism behind the 80–90% token savings. Self-hosted only, AGPL. Requires Docker and an LLM provider for extraction.</p>
|
||||
<div class="pc-tags">
|
||||
<span class="pill pill-green">Self-host, free</span>
|
||||
<span class="pill pill-amber">Self-host only</span>
|
||||
@@ -694,7 +703,7 @@
|
||||
<ul class="picker-picks">
|
||||
<li class="picker-pick">
|
||||
<div class="pick-rank">1</div>
|
||||
<div><span class="pick-name">Hindsight</span> — <span class="pick-why">reflect operation synthesizes across all stored knowledge; highest LongMemEval scores (91–94.6%)</span></div>
|
||||
<div><span class="pick-name">Hindsight</span> — <span class="pick-why">reflect operation synthesizes across all stored knowledge; highest LongMemEval scores (91.4–94.6%)</span></div>
|
||||
</li>
|
||||
<li class="picker-pick">
|
||||
<div class="pick-rank">2</div>
|
||||
@@ -864,7 +873,7 @@
|
||||
<div class="tradeoff-card">
|
||||
<div class="tradeoff-label">LongMemEval benchmark</div>
|
||||
<div class="tradeoff-winner">Hindsight</div>
|
||||
<div class="tradeoff-note">91–94.6% — independently validated, leads by a wide margin</div>
|
||||
<div class="tradeoff-note">91.4–94.6% — independently validated, leads by a wide margin</div>
|
||||
</div>
|
||||
<div class="tradeoff-card">
|
||||
<div class="tradeoff-label">LoCoMo benchmark</div>
|
||||
@@ -899,7 +908,7 @@
|
||||
<div class="tradeoff-card">
|
||||
<div class="tradeoff-label">Token efficiency</div>
|
||||
<div class="tradeoff-winner">OpenViking</div>
|
||||
<div class="tradeoff-note">80–90% token reduction via L0/L1/L2 tiered loading</div>
|
||||
<div class="tradeoff-note">80–90% token savings — loads only needed resolution depth (abstract → overview → full)</div>
|
||||
</div>
|
||||
<div class="tradeoff-card">
|
||||
<div class="tradeoff-label">Tools exposed</div>
|
||||
|
||||
Reference in New Issue
Block a user