You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit e8c11f8
Browse filesBrowse the repository at this point in the historyBrowse files
> **Support continued Engraphis development with Pro.**[Start a 3-day Pro trial](https://api.engraphis.com/account?plan=pro&interval=monthly&utm_source=engraphis&utm_medium=docs&utm_campaign=pro_conversion&utm_content=readme_intro&trial=pro#billing)
28
28
> or [subscribe to Pro](https://api.engraphis.com/account?plan=pro&interval=monthly&utm_source=engraphis&utm_medium=docs&utm_campaign=pro_conversion&utm_content=readme_intro#billing).
29
29
30
+
---
31
+
32
+
## Measured token and context savings
33
+
34
+
<palign="center">
35
+
<imgsrc="docs/images/context-efficiency.png"alt="Dark chart showing Engraphis using 98.21 percent less long-history context, 73.0 percent less retrieved content per question, 73.9 percent fewer tokens in the smallest useful memory, a 55.38 percent smaller memory response, and 47.8 percent less repeated-memory context after consolidation"width="100%">
36
+
<br>
37
+
<sup>Less repeated history means more room for the task, tools, and useful evidence.</sup>
38
+
</p>
39
+
40
+
<details>
41
+
<summary>See benchmark details and reproduce the results</summary>
42
+
43
+
### Controlled before-and-after example
44
+
45
+
| Retrieval mode | Mean returned memory content | Recall@5 |
The chunked mode returns the relevant passage instead of the whole document: **590.4 fewer tokens
51
+
per question**. Under the same model-context budget, that leaves roughly **590 tokens** for task
52
+
instructions or other relevant evidence.
53
+
54
+
### Measurement details and reproducibility
55
+
56
+
The table below records every current token/context efficiency measurement and its counting
57
+
boundary.
58
+
59
+
| What is counted | Comparison | Measured reduction | Quality held constant |
60
+
|---|---|---|---|
61
+
| Cumulative reader context across a 1,986-question LoCoMo diagnostic | Full-history replay: **49,915,394** tokens → Engraphis: **891,857** tokens |**49,023,537 fewer context tokens** (**98.2133% lower**) | Focused retrieval used far less context; uncapped full history retained higher retrieval recall |
62
+
| Retrieved top-5 memory content, averaged per question | Whole documents: **808.8** tokens → structure-aware chunks: **218.4** tokens |**590.4 fewer tokens per question** (**73.0% lower**, about **3.7× smaller**) | Recall@5 **1.000** in both modes across 6 documents and 18 questions |
63
+
| Smallest returned memory that contains the reference evidence | Whole documents: **162.2** tokens → chunks: **42.4** tokens |**119.8 fewer tokens to evidence** (**73.9% lower**, about **3.8× smaller**) | The same 18 questions had a returned evidence-holding memory in both modes |
64
+
| Serialized MCP recall response across 260 timed CodeMem recalls | Full result: **17,172**`engraphis.regex.v1` tokens → compact result: **7,663** tokens |**9,509 response tokens avoided** (**55.38% lower**) | Recall@5, hit@5, and answer-token recall all **1.000**|
65
+
| Repeated-memory consolidation fixture | 12 related episodic memories: **230** tokens → one digest: **120** tokens |**110 tokens removed from the active digest** (**47.8% lower**) | Original memories remain available for provenance and audit |
66
+
| Packed prompt-context usage in the same CodeMem performance fixture | Hard budget: **1,500** tokens; observed mean: **87.73**; observed maximum: **106**| A hard cap prevents a recall from exceeding its configured context budget | This is usage accounting, not a before/after savings comparison |
67
+
68
+
The compact MCP response avoids duplicating full memory bodies when the packed context and source
69
+
list are enough. That can reduce what an agent must inspect or pass onward, but the fixtures do
70
+
**not** measure model-provider charges, end-to-end task time, or customer cost savings.
71
+
72
+
The measures are deliberately separate and **must not be added together**: chunking counts the
73
+
content of retrieved memory records before `ContextPacker`, whereas compact recall counts the
74
+
serialized MCP response returned to a client. “Tokens to evidence” is the size of the smallest
75
+
retrieved memory record holding the reference evidence; it is not latency or end-to-end answer
76
+
accuracy. Chunking creates more focused stored records (24 chunks rather than 6 whole-document
77
+
memories in this fixture), so this is a context-efficiency result, not a storage-reduction claim.
78
+
79
+
Reproduce the quality and token/context measurements without a network connection or API key:
These are small deterministic correctness and efficiency fixtures, not official LoCoMo /
89
+
LongMemEval QA scores or a third-party leaderboard result. Compact-response counts use the exact
90
+
`engraphis.regex.v1` counter; the chunking evaluation uses its documented deterministic
91
+
normalized-character estimator. Chunking measures retrieved memory content, while compact recall
92
+
measures serialized MCP response size. See [`BENCHMARKS.md`](BENCHMARKS.md) for definitions,
93
+
limitations, canonical external-evaluation requirements, and the no-unsupported-claims policy.
94
+
95
+
</details>
96
+
97
+
---
30
98
31
99
## What Engraphis gives an agent
32
100
@@ -130,73 +198,6 @@ chunking. The activity view records outcomes, never keys, prompts, or raw provid
130
198
131
199
---
132
200
133
-
## Measured token and context savings
134
-
135
-
<palign="center">
136
-
<imgsrc="docs/images/context-efficiency.png"alt="Dark chart showing Engraphis using 98.21 percent less long-history context, 73.0 percent less retrieved content per question, 73.9 percent fewer tokens in the smallest useful memory, a 55.38 percent smaller memory response, and 47.8 percent less repeated-memory context after consolidation"width="100%">
137
-
<br>
138
-
<sup>Less repeated history means more room for the task, tools, and useful evidence.</sup>
139
-
</p>
140
-
141
-
<details>
142
-
<summary>See benchmark details and reproduce the results</summary>
143
-
144
-
### Controlled before-and-after example
145
-
146
-
| Retrieval mode | Mean returned memory content | Recall@5 |
The chunked mode returns the relevant passage instead of the whole document: **590.4 fewer tokens
152
-
per question**. Under the same model-context budget, that leaves roughly **590 tokens** for task
153
-
instructions or other relevant evidence.
154
-
155
-
### Measurement details and reproducibility
156
-
157
-
The table below records every current token/context efficiency measurement and its counting
158
-
boundary.
159
-
160
-
| What is counted | Comparison | Measured reduction | Quality held constant |
161
-
|---|---|---|---|
162
-
| Cumulative reader context across a 1,986-question LoCoMo diagnostic | Full-history replay: **49,915,394** tokens → Engraphis: **891,857** tokens |**49,023,537 fewer context tokens** (**98.2133% lower**) | Focused retrieval used far less context; uncapped full history retained higher retrieval recall |
163
-
| Retrieved top-5 memory content, averaged per question | Whole documents: **808.8** tokens → structure-aware chunks: **218.4** tokens |**590.4 fewer tokens per question** (**73.0% lower**, about **3.7× smaller**) | Recall@5 **1.000** in both modes across 6 documents and 18 questions |
164
-
| Smallest returned memory that contains the reference evidence | Whole documents: **162.2** tokens → chunks: **42.4** tokens |**119.8 fewer tokens to evidence** (**73.9% lower**, about **3.8× smaller**) | The same 18 questions had a returned evidence-holding memory in both modes |
165
-
| Serialized MCP recall response across 260 timed CodeMem recalls | Full result: **17,172**`engraphis.regex.v1` tokens → compact result: **7,663** tokens |**9,509 response tokens avoided** (**55.38% lower**) | Recall@5, hit@5, and answer-token recall all **1.000**|
166
-
| Repeated-memory consolidation fixture | 12 related episodic memories: **230** tokens → one digest: **120** tokens |**110 tokens removed from the active digest** (**47.8% lower**) | Original memories remain available for provenance and audit |
167
-
| Packed prompt-context usage in the same CodeMem performance fixture | Hard budget: **1,500** tokens; observed mean: **87.73**; observed maximum: **106**| A hard cap prevents a recall from exceeding its configured context budget | This is usage accounting, not a before/after savings comparison |
168
-
169
-
The compact MCP response avoids duplicating full memory bodies when the packed context and source
170
-
list are enough. That can reduce what an agent must inspect or pass onward, but the fixtures do
171
-
**not** measure model-provider charges, end-to-end task time, or customer cost savings.
172
-
173
-
The measures are deliberately separate and **must not be added together**: chunking counts the
174
-
content of retrieved memory records before `ContextPacker`, whereas compact recall counts the
175
-
serialized MCP response returned to a client. “Tokens to evidence” is the size of the smallest
176
-
retrieved memory record holding the reference evidence; it is not latency or end-to-end answer
177
-
accuracy. Chunking creates more focused stored records (24 chunks rather than 6 whole-document
178
-
memories in this fixture), so this is a context-efficiency result, not a storage-reduction claim.
179
-
180
-
Reproduce the quality and token/context measurements without a network connection or API key:
0 commit comments